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         ocfs2_journal_dirty(handle, bh);
 282
 283 out_commit:
 284         ocfs2_commit_trans(OCFS2_SB(inode->i_sb), handle);
 285 out:
 286         mlog_exit(ret);
 287         return ret;
 288 }
 289
 290 static int ocfs2_set_inode_size(handle_t *handle,
 291                                 struct inode *inode,
 292                                 struct buffer_head *fe_bh,
 293                                 u64 new_i_size)
 294 {
 295         int status;
 296
 297         mlog_entry_void();
 298         i_size_write(inode, new_i_size);
 299         inode->i_blocks = ocfs2_inode_sector_count(inode);
 300         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
 301
 302         status = ocfs2_mark_inode_dirty(handle, inode, fe_bh);
 303         if (status < 0) {
 304                 mlog_errno(status);
 305                 goto bail;
 306         }
 307
 308 bail:
 309         mlog_exit(status);
 310         return status;
 311 }
 312
 313 int ocfs2_simple_size_update(struct inode *inode,
 314                              struct buffer_head *di_bh,
 315                              u64 new_i_size)
 316 {
 317         int ret;
 318         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 319         handle_t *handle = NULL;
 320
 321         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
 322         if (IS_ERR(handle)) {
 323                 ret = PTR_ERR(handle);
 324                 mlog_errno(ret);
 325                 goto out;
 326         }
 327
 328         ret = ocfs2_set_inode_size(handle, inode, di_bh,
 329                                    new_i_size);
 330         if (ret < 0)
 331                 mlog_errno(ret);
 332
 333         ocfs2_commit_trans(osb, handle);
 334 out:
 335         return ret;
 336 }
 337
 338 static int ocfs2_cow_file_pos(struct inode *inode,
 339                               struct buffer_head *fe_bh,
 340                               u64 offset)
 341 {
 342         int status;
 343         u32 phys, cpos = offset >> OCFS2_SB(inode->i_sb)->s_clustersize_bits;
 344         unsigned int num_clusters = 0;
 345         unsigned int ext_flags = 0;
 346
 347         /*
 348          * If the new offset is aligned to the range of the cluster, there is
 349          * no space for ocfs2_zero_range_for_truncate to fill, so no need to
 350          * CoW either.
 351          */
 352         if ((offset & (OCFS2_SB(inode->i_sb)->s_clustersize - 1)) == 0)
 353                 return 0;
 354
 355         status = ocfs2_get_clusters(inode, cpos, &phys,
 356                                     &num_clusters, &ext_flags);
 357         if (status) {
 358                 mlog_errno(status);
 359                 goto out;
 360         }
 361
 362         if (!(ext_flags & OCFS2_EXT_REFCOUNTED))
 363                 goto out;
 364
 365         return ocfs2_refcount_cow(inode, fe_bh, cpos, 1, cpos+1);
 366
 367 out:
 368         return status;
 369 }
 370
 371 static int ocfs2_orphan_for_truncate(struct ocfs2_super *osb,
 372                                      struct inode *inode,
 373                                      struct buffer_head *fe_bh,
 374                                      u64 new_i_size)
 375 {
 376         int status;
 377         handle_t *handle;
 378         struct ocfs2_dinode *di;
 379         u64 cluster_bytes;
 380
 381         mlog_entry_void();
 382
 383         /*
 384          * We need to CoW the cluster contains the offset if it is reflinked
 385          * since we will call ocfs2_zero_range_for_truncate later which will
 386          * write "0" from offset to the end of the cluster.
 387          */
 388         status = ocfs2_cow_file_pos(inode, fe_bh, new_i_size);
 389         if (status) {
 390                 mlog_errno(status);
 391                 return status;
 392         }
 393
 394         /* TODO: This needs to actually orphan the inode in this
 395          * transaction. */
 396
 397         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
 398         if (IS_ERR(handle)) {
 399                 status = PTR_ERR(handle);
 400                 mlog_errno(status);
 401                 goto out;
 402         }
 403
 404         status = ocfs2_journal_access_di(handle, INODE_CACHE(inode), fe_bh,
 405                                          OCFS2_JOURNAL_ACCESS_WRITE);
 406         if (status < 0) {
 407                 mlog_errno(status);
 408                 goto out_commit;
 409         }
 410
 411         /*
 412          * Do this before setting i_size.
 413          */
 414         cluster_bytes = ocfs2_align_bytes_to_clusters(inode->i_sb, new_i_size);
 415         status = ocfs2_zero_range_for_truncate(inode, handle, new_i_size,
 416                                                cluster_bytes);
 417         if (status) {
 418                 mlog_errno(status);
 419                 goto out_commit;
 420         }
 421
 422         i_size_write(inode, new_i_size);
 423         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
 424
 425         di = (struct ocfs2_dinode *) fe_bh->b_data;
 426         di->i_size = cpu_to_le64(new_i_size);
 427         di->i_ctime = di->i_mtime = cpu_to_le64(inode->i_ctime.tv_sec);
 428         di->i_ctime_nsec = di->i_mtime_nsec = cpu_to_le32(inode->i_ctime.tv_nsec);
 429
 430         ocfs2_journal_dirty(handle, fe_bh);
 431
 432 out_commit:
 433         ocfs2_commit_trans(osb, handle);
 434 out:
 435
 436         mlog_exit(status);
 437         return status;
 438 }
 439
 440 static int ocfs2_truncate_file(struct inode *inode,
 441                                struct buffer_head *di_bh,
 442                                u64 new_i_size)
 443 {
 444         int status = 0;
 445         struct ocfs2_dinode *fe = NULL;
 446         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 447
 448         mlog_entry("(inode = %llu, new_i_size = %llu\n",
 449                    (unsigned long long)OCFS2_I(inode)->ip_blkno,
 450                    (unsigned long long)new_i_size);
 451
 452         /* We trust di_bh because it comes from ocfs2_inode_lock(), which
 453          * already validated it */
 454         fe = (struct ocfs2_dinode *) di_bh->b_data;
 455
 456         mlog_bug_on_msg(le64_to_cpu(fe->i_size) != i_size_read(inode),
 457                         "Inode %llu, inode i_size = %lld != di "
 458                         "i_size = %llu, i_flags = 0x%x\n",
 459                         (unsigned long long)OCFS2_I(inode)->ip_blkno,
 460                         i_size_read(inode),
 461                         (unsigned long long)le64_to_cpu(fe->i_size),
 462                         le32_to_cpu(fe->i_flags));
 463
 464         if (new_i_size > le64_to_cpu(fe->i_size)) {
 465                 mlog(0, "asked to truncate file with size (%llu) to size (%llu)!\n",
 466                      (unsigned long long)le64_to_cpu(fe->i_size),
 467                      (unsigned long long)new_i_size);
 468                 status = -EINVAL;
 469                 mlog_errno(status);
 470                 goto bail;
 471         }
 472
 473         mlog(0, "inode %llu, i_size = %llu, new_i_size = %llu\n",
 474              (unsigned long long)le64_to_cpu(fe->i_blkno),
 475              (unsigned long long)le64_to_cpu(fe->i_size),
 476              (unsigned long long)new_i_size);
 477
 478         /* lets handle the simple truncate cases before doing any more
 479          * cluster locking. */
 480         if (new_i_size == le64_to_cpu(fe->i_size))
 481                 goto bail;
 482
 483         down_write(&OCFS2_I(inode)->ip_alloc_sem);
 484
 485         ocfs2_resv_discard(&osb->osb_la_resmap,
 486                            &OCFS2_I(inode)->ip_la_data_resv);
 487
 488         /*
 489          * The inode lock forced other nodes to sync and drop their
 490          * pages, which (correctly) happens even if we have a truncate
 491          * without allocation change - ocfs2 cluster sizes can be much
 492          * greater than page size, so we have to truncate them
 493          * anyway.
 494          */
 495         unmap_mapping_range(inode->i_mapping, new_i_size + PAGE_SIZE - 1, 0, 1);
 496         truncate_inode_pages(inode->i_mapping, new_i_size);
 497
 498         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
 499                 status = ocfs2_truncate_inline(inode, di_bh, new_i_size,
 500                                                i_size_read(inode), 1);
 501                 if (status)
 502                         mlog_errno(status);
 503
 504                 goto bail_unlock_sem;
 505         }
 506
 507         /* alright, we're going to need to do a full blown alloc size
 508          * change. Orphan the inode so that recovery can complete the
 509          * truncate if necessary. This does the task of marking
 510          * i_size. */
 511         status = ocfs2_orphan_for_truncate(osb, inode, di_bh, new_i_size);
 512         if (status < 0) {
 513                 mlog_errno(status);
 514                 goto bail_unlock_sem;
 515         }
 516
 517         status = ocfs2_commit_truncate(osb, inode, di_bh);
 518         if (status < 0) {
 519                 mlog_errno(status);
 520                 goto bail_unlock_sem;
 521         }
 522
 523         /* TODO: orphan dir cleanup here. */
 524 bail_unlock_sem:
 525         up_write(&OCFS2_I(inode)->ip_alloc_sem);
 526
 527 bail:
 528         if (!status && OCFS2_I(inode)->ip_clusters == 0)
 529                 status = ocfs2_try_remove_refcount_tree(inode, di_bh);
 530
 531         mlog_exit(status);
 532         return status;
 533 }
 534
 535 /*
 536  * extend file allocation only here.
 537  * we'll update all the disk stuff, and oip->alloc_size
 538  *
 539  * expect stuff to be locked, a transaction started and enough data /
 540  * metadata reservations in the contexts.
 541  *
 542  * Will return -EAGAIN, and a reason if a restart is needed.
 543  * If passed in, *reason will always be set, even in error.
 544  */
 545 int ocfs2_add_inode_data(struct ocfs2_super *osb,
 546                          struct inode *inode,
 547                          u32 *logical_offset,
 548                          u32 clusters_to_add,
 549                          int mark_unwritten,
 550                          struct buffer_head *fe_bh,
 551                          handle_t *handle,
 552                          struct ocfs2_alloc_context *data_ac,
 553                          struct ocfs2_alloc_context *meta_ac,
 554                          enum ocfs2_alloc_restarted *reason_ret)
 555 {
 556         int ret;
 557         struct ocfs2_extent_tree et;
 558
 559         ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), fe_bh);
 560         ret = ocfs2_add_clusters_in_btree(handle, &et, logical_offset,
 561                                           clusters_to_add, mark_unwritten,
 562                                           data_ac, meta_ac, reason_ret);
 563
 564         return ret;
 565 }
 566
 567 static int __ocfs2_extend_allocation(struct inode *inode, u32 logical_start,
 568                                      u32 clusters_to_add, int mark_unwritten)
 569 {
 570         int status = 0;
 571         int restart_func = 0;
 572         int credits;
 573         u32 prev_clusters;
 574         struct buffer_head *bh = NULL;
 575         struct ocfs2_dinode *fe = NULL;
 576         handle_t *handle = NULL;
 577         struct ocfs2_alloc_context *data_ac = NULL;
 578         struct ocfs2_alloc_context *meta_ac = NULL;
 579         enum ocfs2_alloc_restarted why;
 580         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 581         struct ocfs2_extent_tree et;
 582         int did_quota = 0;
 583
 584         mlog_entry("(clusters_to_add = %u)\n", clusters_to_add);
 585
 586         /*
 587          * This function only exists for file systems which don't
 588          * support holes.
 589          */
 590         BUG_ON(mark_unwritten && !ocfs2_sparse_alloc(osb));
 591
 592         status = ocfs2_read_inode_block(inode, &bh);
 593         if (status < 0) {
 594                 mlog_errno(status);
 595                 goto leave;
 596         }
 597         fe = (struct ocfs2_dinode *) bh->b_data;
 598
 599 restart_all:
 600         BUG_ON(le32_to_cpu(fe->i_clusters) != OCFS2_I(inode)->ip_clusters);
 601
 602         mlog(0, "extend inode %llu, i_size = %lld, di->i_clusters = %u, "
 603              "clusters_to_add = %u\n",
 604              (unsigned long long)OCFS2_I(inode)->ip_blkno,
 605              (long long)i_size_read(inode), le32_to_cpu(fe->i_clusters),
 606              clusters_to_add);
 607         ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), bh);
 608         status = ocfs2_lock_allocators(inode, &et, clusters_to_add, 0,
 609                                        &data_ac, &meta_ac);
 610         if (status) {
 611                 mlog_errno(status);
 612                 goto leave;
 613         }
 614
 615         credits = ocfs2_calc_extend_credits(osb->sb, &fe->id2.i_list,
 616                                             clusters_to_add);
 617         handle = ocfs2_start_trans(osb, credits);
 618         if (IS_ERR(handle)) {
 619                 status = PTR_ERR(handle);
 620                 handle = NULL;
 621                 mlog_errno(status);
 622                 goto leave;
 623         }
 624
 625 restarted_transaction:
 626         status = dquot_alloc_space_nodirty(inode,
 627                         ocfs2_clusters_to_bytes(osb->sb, clusters_to_add));
 628         if (status)
 629                 goto leave;
 630         did_quota = 1;
 631
 632         /* reserve a write to the file entry early on - that we if we
 633          * run out of credits in the allocation path, we can still
 634          * update i_size. */
 635         status = ocfs2_journal_access_di(handle, INODE_CACHE(inode), bh,
 636                                          OCFS2_JOURNAL_ACCESS_WRITE);
 637         if (status < 0) {
 638                 mlog_errno(status);
 639                 goto leave;
 640         }
 641
 642         prev_clusters = OCFS2_I(inode)->ip_clusters;
 643
 644         status = ocfs2_add_inode_data(osb,
 645                                       inode,
 646                                       &logical_start,
 647                                       clusters_to_add,
 648                                       mark_unwritten,
 649                                       bh,
 650                                       handle,
 651                                       data_ac,
 652                                       meta_ac,
 653                                       &why);
 654         if ((status < 0) && (status != -EAGAIN)) {
 655                 if (status != -ENOSPC)
 656                         mlog_errno(status);
 657                 goto leave;
 658         }
 659
 660         ocfs2_journal_dirty(handle, bh);
 661
 662         spin_lock(&OCFS2_I(inode)->ip_lock);
 663         clusters_to_add -= (OCFS2_I(inode)->ip_clusters - prev_clusters);
 664         spin_unlock(&OCFS2_I(inode)->ip_lock);
 665         /* Release unused quota reservation */
 666         dquot_free_space(inode,
 667                         ocfs2_clusters_to_bytes(osb->sb, clusters_to_add));
 668         did_quota = 0;
 669
 670         if (why != RESTART_NONE && clusters_to_add) {
 671                 if (why == RESTART_META) {
 672                         mlog(0, "restarting function.\n");
 673                         restart_func = 1;
 674                         status = 0;
 675                 } else {
 676                         BUG_ON(why != RESTART_TRANS);
 677
 678                         mlog(0, "restarting transaction.\n");
 679                         /* TODO: This can be more intelligent. */
 680                         credits = ocfs2_calc_extend_credits(osb->sb,
 681                                                             &fe->id2.i_list,
 682                                                             clusters_to_add);
 683                         status = ocfs2_extend_trans(handle, credits);
 684                         if (status < 0) {
 685                                 /* handle still has to be committed at
 686                                  * this point. */
 687                                 status = -ENOMEM;
 688                                 mlog_errno(status);
 689                                 goto leave;
 690                         }
 691                         goto restarted_transaction;
 692                 }
 693         }
 694
 695         mlog(0, "fe: i_clusters = %u, i_size=%llu\n",
 696              le32_to_cpu(fe->i_clusters),
 697              (unsigned long long)le64_to_cpu(fe->i_size));
 698         mlog(0, "inode: ip_clusters=%u, i_size=%lld\n",
 699              OCFS2_I(inode)->ip_clusters, (long long)i_size_read(inode));
 700
 701 leave:
 702         if (status < 0 && did_quota)
 703                 dquot_free_space(inode,
 704                         ocfs2_clusters_to_bytes(osb->sb, clusters_to_add));
 705         if (handle) {
 706                 ocfs2_commit_trans(osb, handle);
 707                 handle = NULL;
 708         }
 709         if (data_ac) {
 710                 ocfs2_free_alloc_context(data_ac);
 711                 data_ac = NULL;
 712         }
 713         if (meta_ac) {
 714                 ocfs2_free_alloc_context(meta_ac);
 715                 meta_ac = NULL;
 716         }
 717         if ((!status) && restart_func) {
 718                 restart_func = 0;
 719                 goto restart_all;
 720         }
 721         brelse(bh);
 722         bh = NULL;
 723
 724         mlog_exit(status);
 725         return status;
 726 }
 727
 728 /* Some parts of this taken from generic_cont_expand, which turned out
 729  * to be too fragile to do exactly what we need without us having to
 730  * worry about recursive locking in ->write_begin() and ->write_end(). */
 731 static int ocfs2_write_zero_page(struct inode *inode,
 732                                  u64 size)
 733 {
 734         struct address_space *mapping = inode->i_mapping;
 735         struct page *page;
 736         unsigned long index;
 737         unsigned int offset;
 738         handle_t *handle = NULL;
 739         int ret;
 740
 741         offset = (size & (PAGE_CACHE_SIZE-1)); /* Within page */
 742         /* ugh.  in prepare/commit_write, if from==to==start of block, we
 743         ** skip the prepare.  make sure we never send an offset for the start
 744         ** of a block
 745         */
 746         if ((offset & (inode->i_sb->s_blocksize - 1)) == 0) {
 747                 offset++;
 748         }
 749         index = size >> PAGE_CACHE_SHIFT;
 750
 751         page = grab_cache_page(mapping, index);
 752         if (!page) {
 753                 ret = -ENOMEM;
 754                 mlog_errno(ret);
 755                 goto out;
 756         }
 757
 758         ret = ocfs2_prepare_write_nolock(inode, page, offset, offset);
 759         if (ret < 0) {
 760                 mlog_errno(ret);
 761                 goto out_unlock;
 762         }
 763
 764         if (ocfs2_should_order_data(inode)) {
 765                 handle = ocfs2_start_walk_page_trans(inode, page, offset,
 766                                                      offset);
 767                 if (IS_ERR(handle)) {
 768                         ret = PTR_ERR(handle);
 769                         handle = NULL;
 770                         goto out_unlock;
 771                 }
 772         }
 773
 774         /* must not update i_size! */
 775         ret = block_commit_write(page, offset, offset);
 776         if (ret < 0)
 777                 mlog_errno(ret);
 778         else
 779                 ret = 0;
 780
 781         if (handle)
 782                 ocfs2_commit_trans(OCFS2_SB(inode->i_sb), handle);
 783 out_unlock:
 784         unlock_page(page);
 785         page_cache_release(page);
 786 out:
 787         return ret;
 788 }
 789
 790 static int ocfs2_zero_extend(struct inode *inode,
 791                              u64 zero_to_size)
 792 {
 793         int ret = 0;
 794         u64 start_off;
 795         struct super_block *sb = inode->i_sb;
 796
 797         start_off = ocfs2_align_bytes_to_blocks(sb, i_size_read(inode));
 798         while (start_off < zero_to_size) {
 799                 ret = ocfs2_write_zero_page(inode, start_off);
 800                 if (ret < 0) {
 801                         mlog_errno(ret);
 802                         goto out;
 803                 }
 804
 805                 start_off += sb->s_blocksize;
 806
 807                 /*
 808                  * Very large extends have the potential to lock up
 809                  * the cpu for extended periods of time.
 810                  */
 811                 cond_resched();
 812         }
 813
 814 out:
 815         return ret;
 816 }
 817
 818 int ocfs2_extend_no_holes(struct inode *inode, u64 new_i_size, u64 zero_to)
 819 {
 820         int ret;
 821         u32 clusters_to_add;
 822         struct ocfs2_inode_info *oi = OCFS2_I(inode);
 823
 824         clusters_to_add = ocfs2_clusters_for_bytes(inode->i_sb, new_i_size);
 825         if (clusters_to_add < oi->ip_clusters)
 826                 clusters_to_add = 0;
 827         else
 828                 clusters_to_add -= oi->ip_clusters;
 829
 830         if (clusters_to_add) {
 831                 ret = __ocfs2_extend_allocation(inode, oi->ip_clusters,
 832                                                 clusters_to_add, 0);
 833                 if (ret) {
 834                         mlog_errno(ret);
 835                         goto out;
 836                 }
 837         }
 838
 839         /*
 840          * Call this even if we don't add any clusters to the tree. We
 841          * still need to zero the area between the old i_size and the
 842          * new i_size.
 843          */
 844         ret = ocfs2_zero_extend(inode, zero_to);
 845         if (ret < 0)
 846                 mlog_errno(ret);
 847
 848 out:
 849         return ret;
 850 }
 851
 852 static int ocfs2_extend_file(struct inode *inode,
 853                              struct buffer_head *di_bh,
 854                              u64 new_i_size)
 855 {
 856         int ret = 0;
 857         struct ocfs2_inode_info *oi = OCFS2_I(inode);
 858
 859         BUG_ON(!di_bh);
 860
 861         /* setattr sometimes calls us like this. */
 862         if (new_i_size == 0)
 863                 goto out;
 864
 865         if (i_size_read(inode) == new_i_size)
 866                 goto out;
 867         BUG_ON(new_i_size < i_size_read(inode));
 868
 869         /*
 870          * Fall through for converting inline data, even if the fs
 871          * supports sparse files.
 872          *
 873          * The check for inline data here is legal - nobody can add
 874          * the feature since we have i_mutex. We must check it again
 875          * after acquiring ip_alloc_sem though, as paths like mmap
 876          * might have raced us to converting the inode to extents.
 877          */
 878         if (!(oi->ip_dyn_features & OCFS2_INLINE_DATA_FL)
 879             && ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)))
 880                 goto out_update_size;
 881
 882         /*
 883          * The alloc sem blocks people in read/write from reading our
 884          * allocation until we're done changing it. We depend on
 885          * i_mutex to block other extend/truncate calls while we're
 886          * here.
 887          */
 888         down_write(&oi->ip_alloc_sem);
 889
 890         if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
 891                 /*
 892                  * We can optimize small extends by keeping the inodes
 893                  * inline data.
 894                  */
 895                 if (ocfs2_size_fits_inline_data(di_bh, new_i_size)) {
 896                         up_write(&oi->ip_alloc_sem);
 897                         goto out_update_size;
 898                 }
 899
 900                 ret = ocfs2_convert_inline_data_to_extents(inode, di_bh);
 901                 if (ret) {
 902                         up_write(&oi->ip_alloc_sem);
 903
 904                         mlog_errno(ret);
 905                         goto out;
 906                 }
 907         }
 908
 909         if (!ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)))
 910                 ret = ocfs2_extend_no_holes(inode, new_i_size, new_i_size);
 911
 912         up_write(&oi->ip_alloc_sem);
 913
 914         if (ret < 0) {
 915                 mlog_errno(ret);
 916                 goto out;
 917         }
 918
 919 out_update_size:
 920         ret = ocfs2_simple_size_update(inode, di_bh, new_i_size);
 921         if (ret < 0)
 922                 mlog_errno(ret);
 923
 924 out:
 925         return ret;
 926 }
 927
 928 int ocfs2_setattr(struct dentry *dentry, struct iattr *attr)
 929 {
 930         int status = 0, size_change;
 931         struct inode *inode = dentry->d_inode;
 932         struct super_block *sb = inode->i_sb;
 933         struct ocfs2_super *osb = OCFS2_SB(sb);
 934         struct buffer_head *bh = NULL;
 935         handle_t *handle = NULL;
 936         int qtype;
 937         struct dquot *transfer_from[MAXQUOTAS] = { };
 938         struct dquot *transfer_to[MAXQUOTAS] = { };
 939
 940         mlog_entry("(0x%p, '%.*s')\n", dentry,
 941                    dentry->d_name.len, dentry->d_name.name);
 942
 943         /* ensuring we don't even attempt to truncate a symlink */
 944         if (S_ISLNK(inode->i_mode))
 945                 attr->ia_valid &= ~ATTR_SIZE;
 946
 947         if (attr->ia_valid & ATTR_MODE)
 948                 mlog(0, "mode change: %d\n", attr->ia_mode);
 949         if (attr->ia_valid & ATTR_UID)
 950                 mlog(0, "uid change: %d\n", attr->ia_uid);
 951         if (attr->ia_valid & ATTR_GID)
 952                 mlog(0, "gid change: %d\n", attr->ia_gid);
 953         if (attr->ia_valid & ATTR_SIZE)
 954                 mlog(0, "size change...\n");
 955         if (attr->ia_valid & (ATTR_ATIME | ATTR_MTIME | ATTR_CTIME))
 956                 mlog(0, "time change...\n");
 957
 958 #define OCFS2_VALID_ATTRS (ATTR_ATIME | ATTR_MTIME | ATTR_CTIME | ATTR_SIZE \
 959                            | ATTR_GID | ATTR_UID | ATTR_MODE)
 960         if (!(attr->ia_valid & OCFS2_VALID_ATTRS)) {
 961                 mlog(0, "can't handle attrs: 0x%x\n", attr->ia_valid);
 962                 return 0;
 963         }
 964
 965         status = inode_change_ok(inode, attr);
 966         if (status)
 967                 return status;
 968
 969         size_change = S_ISREG(inode->i_mode) && attr->ia_valid & ATTR_SIZE;
 970         if (size_change) {
 971                 dquot_initialize(inode);
 972
 973                 status = ocfs2_rw_lock(inode, 1);
 974                 if (status < 0) {
 975                         mlog_errno(status);
 976                         goto bail;
 977                 }
 978         }
 979
 980         status = ocfs2_inode_lock(inode, &bh, 1);
 981         if (status < 0) {
 982                 if (status != -ENOENT)
 983                         mlog_errno(status);
 984                 goto bail_unlock_rw;
 985         }
 986
 987         if (size_change && attr->ia_size != i_size_read(inode)) {
 988                 status = inode_newsize_ok(inode, attr->ia_size);
 989                 if (status)
 990                         goto bail_unlock;
 991
 992                 if (i_size_read(inode) > attr->ia_size) {
 993                         if (ocfs2_should_order_data(inode)) {
 994                                 status = ocfs2_begin_ordered_truncate(inode,
 995                                                                       attr->ia_size);
 996                                 if (status)
 997                                         goto bail_unlock;
 998                         }
 999                         status = ocfs2_truncate_file(inode, bh, attr->ia_size);
1000                 } else
1001                         status = ocfs2_extend_file(inode, bh, attr->ia_size);
1002                 if (status < 0) {
1003                         if (status != -ENOSPC)
1004                                 mlog_errno(status);
1005                         status = -ENOSPC;
1006                         goto bail_unlock;
1007                 }
1008         }
1009
1010         if ((attr->ia_valid & ATTR_UID && attr->ia_uid != inode->i_uid) ||
1011             (attr->ia_valid & ATTR_GID && attr->ia_gid != inode->i_gid)) {
1012                 /*
1013                  * Gather pointers to quota structures so that allocation /
1014                  * freeing of quota structures happens here and not inside
1015                  * dquot_transfer() where we have problems with lock ordering
1016                  */
1017                 if (attr->ia_valid & ATTR_UID && attr->ia_uid != inode->i_uid
1018                     && OCFS2_HAS_RO_COMPAT_FEATURE(sb,
1019                     OCFS2_FEATURE_RO_COMPAT_USRQUOTA)) {
1020                         transfer_to[USRQUOTA] = dqget(sb, attr->ia_uid,
1021                                                       USRQUOTA);
1022                         transfer_from[USRQUOTA] = dqget(sb, inode->i_uid,
1023                                                         USRQUOTA);
1024                         if (!transfer_to[USRQUOTA] || !transfer_from[USRQUOTA]) {
1025                                 status = -ESRCH;
1026                                 goto bail_unlock;
1027                         }
1028                 }
1029                 if (attr->ia_valid & ATTR_GID && attr->ia_gid != inode->i_gid
1030                     && OCFS2_HAS_RO_COMPAT_FEATURE(sb,
1031                     OCFS2_FEATURE_RO_COMPAT_GRPQUOTA)) {
1032                         transfer_to[GRPQUOTA] = dqget(sb, attr->ia_gid,
1033                                                       GRPQUOTA);
1034                         transfer_from[GRPQUOTA] = dqget(sb, inode->i_gid,
1035                                                         GRPQUOTA);
1036                         if (!transfer_to[GRPQUOTA] || !transfer_from[GRPQUOTA]) {
1037                                 status = -ESRCH;
1038                                 goto bail_unlock;
1039                         }
1040                 }
1041                 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS +
1042                                            2 * ocfs2_quota_trans_credits(sb));
1043                 if (IS_ERR(handle)) {
1044                         status = PTR_ERR(handle);
1045                         mlog_errno(status);
1046                         goto bail_unlock;
1047                 }
1048                 status = dquot_transfer(inode, attr);
1049                 if (status < 0)
1050                         goto bail_commit;
1051         } else {
1052                 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1053                 if (IS_ERR(handle)) {
1054                         status = PTR_ERR(handle);
1055                         mlog_errno(status);
1056                         goto bail_unlock;
1057                 }
1058         }
1059
1060         /*
1061          * This will intentionally not wind up calling vmtruncate(),
1062          * since all the work for a size change has been done above.
1063          * Otherwise, we could get into problems with truncate as
1064          * ip_alloc_sem is used there to protect against i_size
1065          * changes.
1066          */
1067         status = inode_setattr(inode, attr);
1068         if (status < 0) {
1069                 mlog_errno(status);
1070                 goto bail_commit;
1071         }
1072
1073         status = ocfs2_mark_inode_dirty(handle, inode, bh);
1074         if (status < 0)
1075                 mlog_errno(status);
1076
1077 bail_commit:
1078         ocfs2_commit_trans(osb, handle);
1079 bail_unlock:
1080         ocfs2_inode_unlock(inode, 1);
1081 bail_unlock_rw:
1082         if (size_change)
1083                 ocfs2_rw_unlock(inode, 1);
1084 bail:
1085         brelse(bh);
1086
1087         /* Release quota pointers in case we acquired them */
1088         for (qtype = 0; qtype < MAXQUOTAS; qtype++) {
1089                 dqput(transfer_to[qtype]);
1090                 dqput(transfer_from[qtype]);
1091         }
1092
1093         if (!status && attr->ia_valid & ATTR_MODE) {
1094                 status = ocfs2_acl_chmod(inode);
1095                 if (status < 0)
1096                         mlog_errno(status);
1097         }
1098
1099         mlog_exit(status);
1100         return status;
1101 }
1102
1103 int ocfs2_getattr(struct vfsmount *mnt,
1104                   struct dentry *dentry,
1105                   struct kstat *stat)
1106 {
1107         struct inode *inode = dentry->d_inode;
1108         struct super_block *sb = dentry->d_inode->i_sb;
1109         struct ocfs2_super *osb = sb->s_fs_info;
1110         int err;
1111
1112         mlog_entry_void();
1113
1114         err = ocfs2_inode_revalidate(dentry);
1115         if (err) {
1116                 if (err != -ENOENT)
1117                         mlog_errno(err);
1118                 goto bail;
1119         }
1120
1121         generic_fillattr(inode, stat);
1122
1123         /* We set the blksize from the cluster size for performance */
1124         stat->blksize = osb->s_clustersize;
1125
1126 bail:
1127         mlog_exit(err);
1128
1129         return err;
1130 }
1131
1132 int ocfs2_permission(struct inode *inode, int mask)
1133 {
1134         int ret;
1135
1136         mlog_entry_void();
1137
1138         ret = ocfs2_inode_lock(inode, NULL, 0);
1139         if (ret) {
1140                 if (ret != -ENOENT)
1141                         mlog_errno(ret);
1142                 goto out;
1143         }
1144
1145         ret = generic_permission(inode, mask, ocfs2_check_acl);
1146
1147         ocfs2_inode_unlock(inode, 0);
1148 out:
1149         mlog_exit(ret);
1150         return ret;
1151 }
1152
1153 static int __ocfs2_write_remove_suid(struct inode *inode,
1154                                      struct buffer_head *bh)
1155 {
1156         int ret;
1157         handle_t *handle;
1158         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1159         struct ocfs2_dinode *di;
1160
1161         mlog_entry("(Inode %llu, mode 0%o)\n",
1162                    (unsigned long long)OCFS2_I(inode)->ip_blkno, inode->i_mode);
1163
1164         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1165         if (IS_ERR(handle)) {
1166                 ret = PTR_ERR(handle);
1167                 mlog_errno(ret);
1168                 goto out;
1169         }
1170
1171         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), bh,
1172                                       OCFS2_JOURNAL_ACCESS_WRITE);
1173         if (ret < 0) {
1174                 mlog_errno(ret);
1175                 goto out_trans;
1176         }
1177
1178         inode->i_mode &= ~S_ISUID;
1179         if ((inode->i_mode & S_ISGID) && (inode->i_mode & S_IXGRP))
1180                 inode->i_mode &= ~S_ISGID;
1181
1182         di = (struct ocfs2_dinode *) bh->b_data;
1183         di->i_mode = cpu_to_le16(inode->i_mode);
1184
1185         ocfs2_journal_dirty(handle, bh);
1186
1187 out_trans:
1188         ocfs2_commit_trans(osb, handle);
1189 out:
1190         mlog_exit(ret);
1191         return ret;
1192 }
1193
1194 /*
1195  * Will look for holes and unwritten extents in the range starting at
1196  * pos for count bytes (inclusive).
1197  */
1198 static int ocfs2_check_range_for_holes(struct inode *inode, loff_t pos,
1199                                        size_t count)
1200 {
1201         int ret = 0;
1202         unsigned int extent_flags;
1203         u32 cpos, clusters, extent_len, phys_cpos;
1204         struct super_block *sb = inode->i_sb;
1205
1206         cpos = pos >> OCFS2_SB(sb)->s_clustersize_bits;
1207         clusters = ocfs2_clusters_for_bytes(sb, pos + count) - cpos;
1208
1209         while (clusters) {
1210                 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos, &extent_len,
1211                                          &extent_flags);
1212                 if (ret < 0) {
1213                         mlog_errno(ret);
1214                         goto out;
1215                 }
1216
1217                 if (phys_cpos == 0 || (extent_flags & OCFS2_EXT_UNWRITTEN)) {
1218                         ret = 1;
1219                         break;
1220                 }
1221
1222                 if (extent_len > clusters)
1223                         extent_len = clusters;
1224
1225                 clusters -= extent_len;
1226                 cpos += extent_len;
1227         }
1228 out:
1229         return ret;
1230 }
1231
1232 static int ocfs2_write_remove_suid(struct inode *inode)
1233 {
1234         int ret;
1235         struct buffer_head *bh = NULL;
1236
1237         ret = ocfs2_read_inode_block(inode, &bh);
1238         if (ret < 0) {
1239                 mlog_errno(ret);
1240                 goto out;
1241         }
1242
1243         ret =  __ocfs2_write_remove_suid(inode, bh);
1244 out:
1245         brelse(bh);
1246         return ret;
1247 }
1248
1249 /*
1250  * Allocate enough extents to cover the region starting at byte offset
1251  * start for len bytes. Existing extents are skipped, any extents
1252  * added are marked as "unwritten".
1253  */
1254 static int ocfs2_allocate_unwritten_extents(struct inode *inode,
1255                                             u64 start, u64 len)
1256 {
1257         int ret;
1258         u32 cpos, phys_cpos, clusters, alloc_size;
1259         u64 end = start + len;
1260         struct buffer_head *di_bh = NULL;
1261
1262         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1263                 ret = ocfs2_read_inode_block(inode, &di_bh);
1264                 if (ret) {
1265                         mlog_errno(ret);
1266                         goto out;
1267                 }
1268
1269                 /*
1270                  * Nothing to do if the requested reservation range
1271                  * fits within the inode.
1272                  */
1273                 if (ocfs2_size_fits_inline_data(di_bh, end))
1274                         goto out;
1275
1276                 ret = ocfs2_convert_inline_data_to_extents(inode, di_bh);
1277                 if (ret) {
1278                         mlog_errno(ret);
1279                         goto out;
1280                 }
1281         }
1282
1283         /*
1284          * We consider both start and len to be inclusive.
1285          */
1286         cpos = start >> OCFS2_SB(inode->i_sb)->s_clustersize_bits;
1287         clusters = ocfs2_clusters_for_bytes(inode->i_sb, start + len);
1288         clusters -= cpos;
1289
1290         while (clusters) {
1291                 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos,
1292                                          &alloc_size, NULL);
1293                 if (ret) {
1294                         mlog_errno(ret);
1295                         goto out;
1296                 }
1297
1298                 /*
1299                  * Hole or existing extent len can be arbitrary, so
1300                  * cap it to our own allocation request.
1301                  */
1302                 if (alloc_size > clusters)
1303                         alloc_size = clusters;
1304
1305                 if (phys_cpos) {
1306                         /*
1307                          * We already have an allocation at this
1308                          * region so we can safely skip it.
1309                          */
1310                         goto next;
1311                 }
1312
1313                 ret = __ocfs2_extend_allocation(inode, cpos, alloc_size, 1);
1314                 if (ret) {
1315                         if (ret != -ENOSPC)
1316                                 mlog_errno(ret);
1317                         goto out;
1318                 }
1319
1320 next:
1321                 cpos += alloc_size;
1322                 clusters -= alloc_size;
1323         }
1324
1325         ret = 0;
1326 out:
1327
1328         brelse(di_bh);
1329         return ret;
1330 }
1331
1332 /*
1333  * Truncate a byte range, avoiding pages within partial clusters. This
1334  * preserves those pages for the zeroing code to write to.
1335  */
1336 static void ocfs2_truncate_cluster_pages(struct inode *inode, u64 byte_start,
1337                                          u64 byte_len)
1338 {
1339         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1340         loff_t start, end;
1341         struct address_space *mapping = inode->i_mapping;
1342
1343         start = (loff_t)ocfs2_align_bytes_to_clusters(inode->i_sb, byte_start);
1344         end = byte_start + byte_len;
1345         end = end & ~(osb->s_clustersize - 1);
1346
1347         if (start < end) {
1348                 unmap_mapping_range(mapping, start, end - start, 0);
1349                 truncate_inode_pages_range(mapping, start, end - 1);
1350         }
1351 }
1352
1353 static int ocfs2_zero_partial_clusters(struct inode *inode,
1354                                        u64 start, u64 len)
1355 {
1356         int ret = 0;
1357         u64 tmpend, end = start + len;
1358         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1359         unsigned int csize = osb->s_clustersize;
1360         handle_t *handle;
1361
1362         /*
1363          * The "start" and "end" values are NOT necessarily part of
1364          * the range whose allocation is being deleted. Rather, this
1365          * is what the user passed in with the request. We must zero
1366          * partial clusters here. There's no need to worry about
1367          * physical allocation - the zeroing code knows to skip holes.
1368          */
1369         mlog(0, "byte start: %llu, end: %llu\n",
1370              (unsigned long long)start, (unsigned long long)end);
1371
1372         /*
1373          * If both edges are on a cluster boundary then there's no
1374          * zeroing required as the region is part of the allocation to
1375          * be truncated.
1376          */
1377         if ((start & (csize - 1)) == 0 && (end & (csize - 1)) == 0)
1378                 goto out;
1379
1380         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1381         if (IS_ERR(handle)) {
1382                 ret = PTR_ERR(handle);
1383                 mlog_errno(ret);
1384                 goto out;
1385         }
1386
1387         /*
1388          * We want to get the byte offset of the end of the 1st cluster.
1389          */
1390         tmpend = (u64)osb->s_clustersize + (start & ~(osb->s_clustersize - 1));
1391         if (tmpend > end)
1392                 tmpend = end;
1393
1394         mlog(0, "1st range: start: %llu, tmpend: %llu\n",
1395              (unsigned long long)start, (unsigned long long)tmpend);
1396
1397         ret = ocfs2_zero_range_for_truncate(inode, handle, start, tmpend);
1398         if (ret)
1399                 mlog_errno(ret);
1400
1401         if (tmpend < end) {
1402                 /*
1403                  * This may make start and end equal, but the zeroing
1404                  * code will skip any work in that case so there's no
1405                  * need to catch it up here.
1406                  */
1407                 start = end & ~(osb->s_clustersize - 1);
1408
1409                 mlog(0, "2nd range: start: %llu, end: %llu\n",
1410                      (unsigned long long)start, (unsigned long long)end);
1411
1412                 ret = ocfs2_zero_range_for_truncate(inode, handle, start, end);
1413                 if (ret)
1414                         mlog_errno(ret);
1415         }
1416
1417         ocfs2_commit_trans(osb, handle);
1418 out:
1419         return ret;
1420 }
1421
1422 static int ocfs2_find_rec(struct ocfs2_extent_list *el, u32 pos)
1423 {
1424         int i;
1425         struct ocfs2_extent_rec *rec = NULL;
1426
1427         for (i = le16_to_cpu(el->l_next_free_rec) - 1; i >= 0; i--) {
1428
1429                 rec = &el->l_recs[i];
1430
1431                 if (le32_to_cpu(rec->e_cpos) < pos)
1432                         break;
1433         }
1434
1435         return i;
1436 }
1437
1438 /*
1439  * Helper to calculate the punching pos and length in one run, we handle the
1440  * following three cases in order:
1441  *
1442  * - remove the entire record
1443  * - remove a partial record
1444  * - no record needs to be removed (hole-punching completed)
1445 */
1446 static void ocfs2_calc_trunc_pos(struct inode *inode,
1447                                  struct ocfs2_extent_list *el,
1448                                  struct ocfs2_extent_rec *rec,
1449                                  u32 trunc_start, u32 *trunc_cpos,
1450                                  u32 *trunc_len, u32 *trunc_end,
1451                                  u64 *blkno, int *done)
1452 {
1453         int ret = 0;
1454         u32 coff, range;
1455
1456         range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
1457
1458         if (le32_to_cpu(rec->e_cpos) >= trunc_start) {
1459                 *trunc_cpos = le32_to_cpu(rec->e_cpos);
1460                 /*
1461                  * Skip holes if any.
1462                  */
1463                 if (range < *trunc_end)
1464                         *trunc_end = range;
1465                 *trunc_len = *trunc_end - le32_to_cpu(rec->e_cpos);
1466                 *blkno = le64_to_cpu(rec->e_blkno);
1467                 *trunc_end = le32_to_cpu(rec->e_cpos);
1468         } else if (range > trunc_start) {
1469                 *trunc_cpos = trunc_start;
1470                 *trunc_len = *trunc_end - trunc_start;
1471                 coff = trunc_start - le32_to_cpu(rec->e_cpos);
1472                 *blkno = le64_to_cpu(rec->e_blkno) +
1473                                 ocfs2_clusters_to_blocks(inode->i_sb, coff);
1474                 *trunc_end = trunc_start;
1475         } else {
1476                 /*
1477                  * It may have two following possibilities:
1478                  *
1479                  * - last record has been removed
1480                  * - trunc_start was within a hole
1481                  *
1482                  * both two cases mean the completion of hole punching.
1483                  */
1484                 ret = 1;
1485         }
1486
1487         *done = ret;
1488 }
1489
1490 static int ocfs2_remove_inode_range(struct inode *inode,
1491                                     struct buffer_head *di_bh, u64 byte_start,
1492                                     u64 byte_len)
1493 {
1494         int ret = 0, flags = 0, done = 0, i;
1495         u32 trunc_start, trunc_len, trunc_end, trunc_cpos, phys_cpos;
1496         u32 cluster_in_el;
1497         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1498         struct ocfs2_cached_dealloc_ctxt dealloc;
1499         struct address_space *mapping = inode->i_mapping;
1500         struct ocfs2_extent_tree et;
1501         struct ocfs2_path *path = NULL;
1502         struct ocfs2_extent_list *el = NULL;
1503         struct ocfs2_extent_rec *rec = NULL;
1504         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
1505         u64 blkno, refcount_loc = le64_to_cpu(di->i_refcount_loc);
1506
1507         ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), di_bh);
1508         ocfs2_init_dealloc_ctxt(&dealloc);
1509
1510         if (byte_len == 0)
1511                 return 0;
1512
1513         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1514                 ret = ocfs2_truncate_inline(inode, di_bh, byte_start,
1515                                             byte_start + byte_len, 0);
1516                 if (ret) {
1517                         mlog_errno(ret);
1518                         goto out;
1519                 }
1520                 /*
1521                  * There's no need to get fancy with the page cache
1522                  * truncate of an inline-data inode. We're talking
1523                  * about less than a page here, which will be cached
1524                  * in the dinode buffer anyway.
1525                  */
1526                 unmap_mapping_range(mapping, 0, 0, 0);
1527                 truncate_inode_pages(mapping, 0);
1528                 goto out;
1529         }
1530
1531         /*
1532          * For reflinks, we may need to CoW 2 clusters which might be
1533          * partially zero'd later, if hole's start and end offset were
1534          * within one cluster(means is not exactly aligned to clustersize).
1535          */
1536
1537         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL) {
1538
1539                 ret = ocfs2_cow_file_pos(inode, di_bh, byte_start);
1540                 if (ret) {
1541                         mlog_errno(ret);
1542                         goto out;
1543                 }
1544
1545                 ret = ocfs2_cow_file_pos(inode, di_bh, byte_start + byte_len);
1546                 if (ret) {
1547                         mlog_errno(ret);
1548                         goto out;
1549                 }
1550         }
1551
1552         trunc_start = ocfs2_clusters_for_bytes(osb->sb, byte_start);
1553         trunc_end = (byte_start + byte_len) >> osb->s_clustersize_bits;
1554         cluster_in_el = trunc_end;
1555
1556         mlog(0, "Inode: %llu, start: %llu, len: %llu, cstart: %u, cend: %u\n",
1557              (unsigned long long)OCFS2_I(inode)->ip_blkno,
1558              (unsigned long long)byte_start,
1559              (unsigned long long)byte_len, trunc_start, trunc_end);
1560
1561         ret = ocfs2_zero_partial_clusters(inode, byte_start, byte_len);
1562         if (ret) {
1563                 mlog_errno(ret);
1564                 goto out;
1565         }
1566
1567         path = ocfs2_new_path_from_et(&et);
1568         if (!path) {
1569                 ret = -ENOMEM;
1570                 mlog_errno(ret);
1571                 goto out;
1572         }
1573
1574         while (trunc_end > trunc_start) {
1575
1576                 ret = ocfs2_find_path(INODE_CACHE(inode), path,
1577                                       cluster_in_el);
1578                 if (ret) {
1579                         mlog_errno(ret);
1580                         goto out;
1581                 }
1582
1583                 el = path_leaf_el(path);
1584
1585                 i = ocfs2_find_rec(el, trunc_end);
1586                 /*
1587                  * Need to go to previous extent block.
1588                  */
1589                 if (i < 0) {
1590                         if (path->p_tree_depth == 0)
1591                                 break;
1592
1593                         ret = ocfs2_find_cpos_for_left_leaf(inode->i_sb,
1594                                                             path,
1595                                                             &cluster_in_el);
1596                         if (ret) {
1597                                 mlog_errno(ret);
1598                                 goto out;
1599                         }
1600
1601                         /*
1602                          * We've reached the leftmost extent block,
1603                          * it's safe to leave.
1604                          */
1605                         if (cluster_in_el == 0)
1606                                 break;
1607
1608                         /*
1609                          * The 'pos' searched for previous extent block is
1610                          * always one cluster less than actual trunc_end.
1611                          */
1612                         trunc_end = cluster_in_el + 1;
1613
1614                         ocfs2_reinit_path(path, 1);
1615
1616                         continue;
1617
1618                 } else
1619                         rec = &el->l_recs[i];
1620
1621                 ocfs2_calc_trunc_pos(inode, el, rec, trunc_start, &trunc_cpos,
1622                                      &trunc_len, &trunc_end, &blkno, &done);
1623                 if (done)
1624                         break;
1625
1626                 flags = rec->e_flags;
1627                 phys_cpos = ocfs2_blocks_to_clusters(inode->i_sb, blkno);
1628
1629                 ret = ocfs2_remove_btree_range(inode, &et, trunc_cpos,
1630                                                phys_cpos, trunc_len, flags,
1631                                                &dealloc, refcount_loc);
1632                 if (ret < 0) {
1633                         mlog_errno(ret);
1634                         goto out;
1635                 }
1636
1637                 cluster_in_el = trunc_end;
1638
1639                 ocfs2_reinit_path(path, 1);
1640         }
1641
1642         ocfs2_truncate_cluster_pages(inode, byte_start, byte_len);
1643
1644 out:
1645         ocfs2_schedule_truncate_log_flush(osb, 1);
1646         ocfs2_run_deallocs(osb, &dealloc);
1647
1648         return ret;
1649 }
1650
1651 /*
1652  * Parts of this function taken from xfs_change_file_space()
1653  */
1654 static int __ocfs2_change_file_space(struct file *file, struct inode *inode,
1655                                      loff_t f_pos, unsigned int cmd,
1656                                      struct ocfs2_space_resv *sr,
1657                                      int change_size)
1658 {
1659         int ret;
1660         s64 llen;
1661         loff_t size;
1662         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1663         struct buffer_head *di_bh = NULL;
1664         handle_t *handle;
1665         unsigned long long max_off = inode->i_sb->s_maxbytes;
1666
1667         if (ocfs2_is_hard_readonly(osb) || ocfs2_is_soft_readonly(osb))
1668                 return -EROFS;
1669
1670         mutex_lock(&inode->i_mutex);
1671
1672         /*
1673          * This prevents concurrent writes on other nodes
1674          */
1675         ret = ocfs2_rw_lock(inode, 1);
1676         if (ret) {
1677                 mlog_errno(ret);
1678                 goto out;
1679         }
1680
1681         ret = ocfs2_inode_lock(inode, &di_bh, 1);
1682         if (ret) {
1683                 mlog_errno(ret);
1684                 goto out_rw_unlock;
1685         }
1686
1687         if (inode->i_flags & (S_IMMUTABLE|S_APPEND)) {
1688                 ret = -EPERM;
1689                 goto out_inode_unlock;
1690         }
1691
1692         switch (sr->l_whence) {
1693         case 0: /*SEEK_SET*/
1694                 break;
1695         case 1: /*SEEK_CUR*/
1696                 sr->l_start += f_pos;
1697                 break;
1698         case 2: /*SEEK_END*/
1699                 sr->l_start += i_size_read(inode);
1700                 break;
1701         default:
1702                 ret = -EINVAL;
1703                 goto out_inode_unlock;
1704         }
1705         sr->l_whence = 0;
1706
1707         llen = sr->l_len > 0 ? sr->l_len - 1 : sr->l_len;
1708
1709         if (sr->l_start < 0
1710             || sr->l_start > max_off
1711             || (sr->l_start + llen) < 0
1712             || (sr->l_start + llen) > max_off) {
1713                 ret = -EINVAL;
1714                 goto out_inode_unlock;
1715         }
1716         size = sr->l_start + sr->l_len;
1717
1718         if (cmd == OCFS2_IOC_RESVSP || cmd == OCFS2_IOC_RESVSP64) {
1719                 if (sr->l_len <= 0) {
1720                         ret = -EINVAL;
1721                         goto out_inode_unlock;
1722                 }
1723         }
1724
1725         if (file && should_remove_suid(file->f_path.dentry)) {
1726                 ret = __ocfs2_write_remove_suid(inode, di_bh);
1727                 if (ret) {
1728                         mlog_errno(ret);
1729                         goto out_inode_unlock;
1730                 }
1731         }
1732
1733         down_write(&OCFS2_I(inode)->ip_alloc_sem);
1734         switch (cmd) {
1735         case OCFS2_IOC_RESVSP:
1736         case OCFS2_IOC_RESVSP64:
1737                 /*
1738                  * This takes unsigned offsets, but the signed ones we
1739                  * pass have been checked against overflow above.
1740                  */
1741                 ret = ocfs2_allocate_unwritten_extents(inode, sr->l_start,
1742                                                        sr->l_len);
1743                 break;
1744         case OCFS2_IOC_UNRESVSP:
1745         case OCFS2_IOC_UNRESVSP64:
1746                 ret = ocfs2_remove_inode_range(inode, di_bh, sr->l_start,
1747                                                sr->l_len);
1748                 break;
1749         default:
1750                 ret = -EINVAL;
1751         }
1752         up_write(&OCFS2_I(inode)->ip_alloc_sem);
1753         if (ret) {
1754                 mlog_errno(ret);
1755                 goto out_inode_unlock;
1756         }
1757
1758         /*
1759          * We update c/mtime for these changes
1760          */
1761         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1762         if (IS_ERR(handle)) {
1763                 ret = PTR_ERR(handle);
1764                 mlog_errno(ret);
1765                 goto out_inode_unlock;
1766         }
1767
1768         if (change_size && i_size_read(inode) < size)
1769                 i_size_write(inode, size);
1770
1771         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
1772         ret = ocfs2_mark_inode_dirty(handle, inode, di_bh);
1773         if (ret < 0)
1774                 mlog_errno(ret);
1775
1776         ocfs2_commit_trans(osb, handle);
1777
1778 out_inode_unlock:
1779         brelse(di_bh);
1780         ocfs2_inode_unlock(inode, 1);
1781 out_rw_unlock:
1782         ocfs2_rw_unlock(inode, 1);
1783
1784 out:
1785         mutex_unlock(&inode->i_mutex);
1786         return ret;
1787 }
1788
1789 int ocfs2_change_file_space(struct file *file, unsigned int cmd,
1790                             struct ocfs2_space_resv *sr)
1791 {
1792         struct inode *inode = file->f_path.dentry->d_inode;
1793         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1794
1795         if ((cmd == OCFS2_IOC_RESVSP || cmd == OCFS2_IOC_RESVSP64) &&
1796             !ocfs2_writes_unwritten_extents(osb))
1797                 return -ENOTTY;
1798         else if ((cmd == OCFS2_IOC_UNRESVSP || cmd == OCFS2_IOC_UNRESVSP64) &&
1799                  !ocfs2_sparse_alloc(osb))
1800                 return -ENOTTY;
1801
1802         if (!S_ISREG(inode->i_mode))
1803                 return -EINVAL;
1804
1805         if (!(file->f_mode & FMODE_WRITE))
1806                 return -EBADF;
1807
1808         return __ocfs2_change_file_space(file, inode, file->f_pos, cmd, sr, 0);
1809 }
1810
1811 static long ocfs2_fallocate(struct inode *inode, int mode, loff_t offset,
1812                             loff_t len)
1813 {
1814         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1815         struct ocfs2_space_resv sr;
1816         int change_size = 1;
1817
1818         if (!ocfs2_writes_unwritten_extents(osb))
1819                 return -EOPNOTSUPP;
1820
1821         if (S_ISDIR(inode->i_mode))
1822                 return -ENODEV;
1823
1824         if (mode & FALLOC_FL_KEEP_SIZE)
1825                 change_size = 0;
1826
1827         sr.l_whence = 0;
1828         sr.l_start = (s64)offset;
1829         sr.l_len = (s64)len;
1830
1831         return __ocfs2_change_file_space(NULL, inode, offset,
1832                                          OCFS2_IOC_RESVSP64, &sr, change_size);
1833 }
1834
1835 int ocfs2_check_range_for_refcount(struct inode *inode, loff_t pos,
1836                                    size_t count)
1837 {
1838         int ret = 0;
1839         unsigned int extent_flags;
1840         u32 cpos, clusters, extent_len, phys_cpos;
1841         struct super_block *sb = inode->i_sb;
1842
1843         if (!ocfs2_refcount_tree(OCFS2_SB(inode->i_sb)) ||
1844             !(OCFS2_I(inode)->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL) ||
1845             OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
1846                 return 0;
1847
1848         cpos = pos >> OCFS2_SB(sb)->s_clustersize_bits;
1849         clusters = ocfs2_clusters_for_bytes(sb, pos + count) - cpos;
1850
1851         while (clusters) {
1852                 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos, &extent_len,
1853                                          &extent_flags);
1854                 if (ret < 0) {
1855                         mlog_errno(ret);
1856                         goto out;
1857                 }
1858
1859                 if (phys_cpos && (extent_flags & OCFS2_EXT_REFCOUNTED)) {
1860                         ret = 1;
1861                         break;
1862                 }
1863
1864                 if (extent_len > clusters)
1865                         extent_len = clusters;
1866
1867                 clusters -= extent_len;
1868                 cpos += extent_len;
1869         }
1870 out:
1871         return ret;
1872 }
1873
1874 static int ocfs2_prepare_inode_for_refcount(struct inode *inode,
1875                                             loff_t pos, size_t count,
1876                                             int *meta_level)
1877 {
1878         int ret;
1879         struct buffer_head *di_bh = NULL;
1880         u32 cpos = pos >> OCFS2_SB(inode->i_sb)->s_clustersize_bits;
1881         u32 clusters =
1882                 ocfs2_clusters_for_bytes(inode->i_sb, pos + count) - cpos;
1883
1884         ret = ocfs2_inode_lock(inode, &di_bh, 1);
1885         if (ret) {
1886                 mlog_errno(ret);
1887                 goto out;
1888         }
1889
1890         *meta_level = 1;
1891
1892         ret = ocfs2_refcount_cow(inode, di_bh, cpos, clusters, UINT_MAX);
1893         if (ret)
1894                 mlog_errno(ret);
1895 out:
1896         brelse(di_bh);
1897         return ret;
1898 }
1899
1900 static int ocfs2_prepare_inode_for_write(struct dentry *dentry,
1901                                          loff_t *ppos,
1902                                          size_t count,
1903                                          int appending,
1904                                          int *direct_io,
1905                                          int *has_refcount)
1906 {
1907         int ret = 0, meta_level = 0;
1908         struct inode *inode = dentry->d_inode;
1909         loff_t saved_pos, end;
1910
1911         /*
1912          * We start with a read level meta lock and only jump to an ex
1913          * if we need to make modifications here.
1914          */
1915         for(;;) {
1916                 ret = ocfs2_inode_lock(inode, NULL, meta_level);
1917                 if (ret < 0) {
1918                         meta_level = -1;
1919                         mlog_errno(ret);
1920                         goto out;
1921                 }
1922
1923                 /* Clear suid / sgid if necessary. We do this here
1924                  * instead of later in the write path because
1925                  * remove_suid() calls ->setattr without any hint that
1926                  * we may have already done our cluster locking. Since
1927                  * ocfs2_setattr() *must* take cluster locks to
1928                  * proceeed, this will lead us to recursively lock the
1929                  * inode. There's also the dinode i_size state which
1930                  * can be lost via setattr during extending writes (we
1931                  * set inode->i_size at the end of a write. */
1932                 if (should_remove_suid(dentry)) {
1933                         if (meta_level == 0) {
1934                                 ocfs2_inode_unlock(inode, meta_level);
1935                                 meta_level = 1;
1936                                 continue;
1937                         }
1938
1939                         ret = ocfs2_write_remove_suid(inode);
1940                         if (ret < 0) {
1941                                 mlog_errno(ret);
1942                                 goto out_unlock;
1943                         }
1944                 }
1945
1946                 /* work on a copy of ppos until we're sure that we won't have
1947                  * to recalculate it due to relocking. */
1948                 if (appending) {
1949                         saved_pos = i_size_read(inode);
1950                         mlog(0, "O_APPEND: inode->i_size=%llu\n", saved_pos);
1951                 } else {
1952                         saved_pos = *ppos;
1953                 }
1954
1955                 end = saved_pos + count;
1956
1957                 ret = ocfs2_check_range_for_refcount(inode, saved_pos, count);
1958                 if (ret == 1) {
1959                         ocfs2_inode_unlock(inode, meta_level);
1960                         meta_level = -1;
1961
1962                         ret = ocfs2_prepare_inode_for_refcount(inode,
1963                                                                saved_pos,
1964                                                                count,
1965                                                                &meta_level);
1966                         if (has_refcount)
1967                                 *has_refcount = 1;
1968                         if (direct_io)
1969                                 *direct_io = 0;
1970                 }
1971
1972                 if (ret < 0) {
1973                         mlog_errno(ret);
1974                         goto out_unlock;
1975                 }
1976
1977                 /*
1978                  * Skip the O_DIRECT checks if we don't need
1979                  * them.
1980                  */
1981                 if (!direct_io || !(*direct_io))
1982                         break;
1983
1984                 /*
1985                  * There's no sane way to do direct writes to an inode
1986                  * with inline data.
1987                  */
1988                 if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1989                         *direct_io = 0;
1990                         break;
1991                 }
1992
1993                 /*
1994                  * Allowing concurrent direct writes means
1995                  * i_size changes wouldn't be synchronized, so
1996                  * one node could wind up truncating another
1997                  * nodes writes.
1998                  */
1999                 if (end > i_size_read(inode)) {
2000                         *direct_io = 0;
2001                         break;
2002                 }
2003
2004                 /*
2005                  * We don't fill holes during direct io, so
2006                  * check for them here. If any are found, the
2007                  * caller will have to retake some cluster
2008                  * locks and initiate the io as buffered.
2009                  */
2010                 ret = ocfs2_check_range_for_holes(inode, saved_pos, count);
2011                 if (ret == 1) {
2012                         *direct_io = 0;
2013                         ret = 0;
2014                 } else if (ret < 0)
2015                         mlog_errno(ret);
2016                 break;
2017         }
2018
2019         if (appending)
2020                 *ppos = saved_pos;
2021
2022 out_unlock:
2023         if (meta_level >= 0)
2024                 ocfs2_inode_unlock(inode, meta_level);
2025
2026 out:
2027         return ret;
2028 }
2029
2030 static ssize_t ocfs2_file_aio_write(struct kiocb *iocb,
2031                                     const struct iovec *iov,
2032                                     unsigned long nr_segs,
2033                                     loff_t pos)
2034 {
2035         int ret, direct_io, appending, rw_level, have_alloc_sem  = 0;
2036         int can_do_direct, has_refcount = 0;
2037         ssize_t written = 0;
2038         size_t ocount;          /* original count */
2039         size_t count;           /* after file limit checks */
2040         loff_t old_size, *ppos = &iocb->ki_pos;
2041         u32 old_clusters;
2042         struct file *file = iocb->ki_filp;
2043         struct inode *inode = file->f_path.dentry->d_inode;
2044         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
2045
2046         mlog_entry("(0x%p, %u, '%.*s')\n", file,
2047                    (unsigned int)nr_segs,
2048                    file->f_path.dentry->d_name.len,
2049                    file->f_path.dentry->d_name.name);
2050
2051         if (iocb->ki_left == 0)
2052                 return 0;
2053
2054         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
2055
2056         appending = file->f_flags & O_APPEND ? 1 : 0;
2057         direct_io = file->f_flags & O_DIRECT ? 1 : 0;
2058
2059         mutex_lock(&inode->i_mutex);
2060
2061 relock:
2062         /* to match setattr's i_mutex -> i_alloc_sem -> rw_lock ordering */
2063         if (direct_io) {
2064                 down_read(&inode->i_alloc_sem);
2065                 have_alloc_sem = 1;
2066         }
2067
2068         /* concurrent O_DIRECT writes are allowed */
2069         rw_level = !direct_io;
2070         ret = ocfs2_rw_lock(inode, rw_level);
2071         if (ret < 0) {
2072                 mlog_errno(ret);
2073                 goto out_sems;
2074         }
2075
2076         can_do_direct = direct_io;
2077         ret = ocfs2_prepare_inode_for_write(file->f_path.dentry, ppos,
2078                                             iocb->ki_left, appending,
2079                                             &can_do_direct, &has_refcount);
2080         if (ret < 0) {
2081                 mlog_errno(ret);
2082                 goto out;
2083         }
2084
2085         /*
2086          * We can't complete the direct I/O as requested, fall back to
2087          * buffered I/O.
2088          */
2089         if (direct_io && !can_do_direct) {
2090                 ocfs2_rw_unlock(inode, rw_level);
2091                 up_read(&inode->i_alloc_sem);
2092
2093                 have_alloc_sem = 0;
2094                 rw_level = -1;
2095
2096                 direct_io = 0;
2097                 goto relock;
2098         }
2099
2100         /*
2101          * To later detect whether a journal commit for sync writes is
2102          * necessary, we sample i_size, and cluster count here.
2103          */
2104         old_size = i_size_read(inode);
2105         old_clusters = OCFS2_I(inode)->ip_clusters;
2106
2107         /* communicate with ocfs2_dio_end_io */
2108         ocfs2_iocb_set_rw_locked(iocb, rw_level);
2109
2110         ret = generic_segment_checks(iov, &nr_segs, &ocount,
2111                                      VERIFY_READ);
2112         if (ret)
2113                 goto out_dio;
2114
2115         count = ocount;
2116         ret = generic_write_checks(file, ppos, &count,
2117                                    S_ISBLK(inode->i_mode));
2118         if (ret)
2119                 goto out_dio;
2120
2121         if (direct_io) {
2122                 written = generic_file_direct_write(iocb, iov, &nr_segs, *ppos,
2123                                                     ppos, count, ocount);
2124                 if (written < 0) {
2125                         /*
2126                          * direct write may have instantiated a few
2127                          * blocks outside i_size. Trim these off again.
2128                          * Don't need i_size_read because we hold i_mutex.
2129                          */
2130                         if (*ppos + count > inode->i_size)
2131                                 vmtruncate(inode, inode->i_size);
2132                         ret = written;
2133                         goto out_dio;
2134                 }
2135         } else {
2136                 current->backing_dev_info = file->f_mapping->backing_dev_info;
2137                 written = generic_file_buffered_write(iocb, iov, nr_segs, *ppos,
2138                                                       ppos, count, 0);
2139                 current->backing_dev_info = NULL;
2140         }
2141
2142 out_dio:
2143         /* buffered aio wouldn't have proper lock coverage today */
2144         BUG_ON(ret == -EIOCBQUEUED && !(file->f_flags & O_DIRECT));
2145
2146         if (((file->f_flags & O_DSYNC) && !direct_io) || IS_SYNC(inode) ||
2147             ((file->f_flags & O_DIRECT) && has_refcount)) {
2148                 ret = filemap_fdatawrite_range(file->f_mapping, pos,
2149                                                pos + count - 1);
2150                 if (ret < 0)
2151                         written = ret;
2152
2153                 if (!ret && ((old_size != i_size_read(inode)) ||
2154                              (old_clusters != OCFS2_I(inode)->ip_clusters) ||
2155                              has_refcount)) {
2156                         ret = jbd2_journal_force_commit(osb->journal->j_journal);
2157                         if (ret < 0)
2158                                 written = ret;
2159                 }
2160
2161                 if (!ret)
2162                         ret = filemap_fdatawait_range(file->f_mapping, pos,
2163                                                       pos + count - 1);
2164         }
2165
2166         /*
2167          * deep in g_f_a_w_n()->ocfs2_direct_IO we pass in a ocfs2_dio_end_io
2168          * function pointer which is called when o_direct io completes so that
2169          * it can unlock our rw lock.  (it's the clustered equivalent of
2170          * i_alloc_sem; protects truncate from racing with pending ios).
2171          * Unfortunately there are error cases which call end_io and others
2172          * that don't.  so we don't have to unlock the rw_lock if either an
2173          * async dio is going to do it in the future or an end_io after an
2174          * error has already done it.
2175          */
2176         if ((ret == -EIOCBQUEUED) || (!ocfs2_iocb_is_rw_locked(iocb))) {
2177                 rw_level = -1;
2178                 have_alloc_sem = 0;
2179         }
2180
2181 out:
2182         if (rw_level != -1)
2183                 ocfs2_rw_unlock(inode, rw_level);
2184
2185 out_sems:
2186         if (have_alloc_sem)
2187                 up_read(&inode->i_alloc_sem);
2188
2189         mutex_unlock(&inode->i_mutex);
2190
2191         if (written)
2192                 ret = written;
2193         mlog_exit(ret);
2194         return ret;
2195 }
2196
2197 static int ocfs2_splice_to_file(struct pipe_inode_info *pipe,
2198                                 struct file *out,
2199                                 struct splice_desc *sd)
2200 {
2201         int ret;
2202
2203         ret = ocfs2_prepare_inode_for_write(out->f_path.dentry, &sd->pos,
2204                                             sd->total_len, 0, NULL, NULL);
2205         if (ret < 0) {
2206                 mlog_errno(ret);
2207                 return ret;
2208         }
2209
2210         return splice_from_pipe_feed(pipe, sd, pipe_to_file);
2211 }
2212
2213 static ssize_t ocfs2_file_splice_write(struct pipe_inode_info *pipe,
2214                                        struct file *out,
2215                                        loff_t *ppos,
2216                                        size_t len,
2217                                        unsigned int flags)
2218 {
2219         int ret;
2220         struct address_space *mapping = out->f_mapping;
2221         struct inode *inode = mapping->host;
2222         struct splice_desc sd = {
2223                 .total_len = len,
2224                 .flags = flags,
2225                 .pos = *ppos,
2226                 .u.file = out,
2227         };
2228
2229         mlog_entry("(0x%p, 0x%p, %u, '%.*s')\n", out, pipe,
2230                    (unsigned int)len,
2231                    out->f_path.dentry->d_name.len,
2232                    out->f_path.dentry->d_name.name);
2233
2234         if (pipe->inode)
2235                 mutex_lock_nested(&pipe->inode->i_mutex, I_MUTEX_PARENT);
2236
2237         splice_from_pipe_begin(&sd);
2238         do {
2239                 ret = splice_from_pipe_next(pipe, &sd);
2240                 if (ret <= 0)
2241                         break;
2242
2243                 mutex_lock_nested(&inode->i_mutex, I_MUTEX_CHILD);
2244                 ret = ocfs2_rw_lock(inode, 1);
2245                 if (ret < 0)
2246                         mlog_errno(ret);
2247                 else {
2248                         ret = ocfs2_splice_to_file(pipe, out, &sd);
2249                         ocfs2_rw_unlock(inode, 1);
2250                 }
2251                 mutex_unlock(&inode->i_mutex);
2252         } while (ret > 0);
2253         splice_from_pipe_end(pipe, &sd);
2254
2255         if (pipe->inode)
2256                 mutex_unlock(&pipe->inode->i_mutex);
2257
2258         if (sd.num_spliced)
2259                 ret = sd.num_spliced;
2260
2261         if (ret > 0) {
2262                 unsigned long nr_pages;
2263                 int err;
2264
2265                 nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
2266
2267                 err = generic_write_sync(out, *ppos, ret);
2268                 if (err)
2269                         ret = err;
2270                 else
2271                         *ppos += ret;
2272
2273                 balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
2274         }
2275
2276         mlog_exit(ret);
2277         return ret;
2278 }
2279
2280 static ssize_t ocfs2_file_splice_read(struct file *in,
2281                                       loff_t *ppos,
2282                                       struct pipe_inode_info *pipe,
2283                                       size_t len,
2284                                       unsigned int flags)
2285 {
2286         int ret = 0, lock_level = 0;
2287         struct inode *inode = in->f_path.dentry->d_inode;
2288
2289         mlog_entry("(0x%p, 0x%p, %u, '%.*s')\n", in, pipe,
2290                    (unsigned int)len,
2291                    in->f_path.dentry->d_name.len,
2292                    in->f_path.dentry->d_name.name);
2293
2294         /*
2295          * See the comment in ocfs2_file_aio_read()
2296          */
2297         ret = ocfs2_inode_lock_atime(inode, in->f_vfsmnt, &lock_level);
2298         if (ret < 0) {
2299                 mlog_errno(ret);
2300                 goto bail;
2301         }
2302         ocfs2_inode_unlock(inode, lock_level);
2303
2304         ret = generic_file_splice_read(in, ppos, pipe, len, flags);
2305
2306 bail:
2307         mlog_exit(ret);
2308         return ret;
2309 }
2310
2311 static ssize_t ocfs2_file_aio_read(struct kiocb *iocb,
2312                                    const struct iovec *iov,
2313                                    unsigned long nr_segs,
2314                                    loff_t pos)
2315 {
2316         int ret = 0, rw_level = -1, have_alloc_sem = 0, lock_level = 0;
2317         struct file *filp = iocb->ki_filp;
2318         struct inode *inode = filp->f_path.dentry->d_inode;
2319
2320         mlog_entry("(0x%p, %u, '%.*s')\n", filp,
2321                    (unsigned int)nr_segs,
2322                    filp->f_path.dentry->d_name.len,
2323                    filp->f_path.dentry->d_name.name);
2324
2325         if (!inode) {
2326                 ret = -EINVAL;
2327                 mlog_errno(ret);
2328                 goto bail;
2329         }
2330
2331         /*
2332          * buffered reads protect themselves in ->readpage().  O_DIRECT reads
2333          * need locks to protect pending reads from racing with truncate.
2334          */
2335         if (filp->f_flags & O_DIRECT) {
2336                 down_read(&inode->i_alloc_sem);
2337                 have_alloc_sem = 1;
2338
2339                 ret = ocfs2_rw_lock(inode, 0);
2340                 if (ret < 0) {
2341                         mlog_errno(ret);
2342                         goto bail;
2343                 }
2344                 rw_level = 0;
2345                 /* communicate with ocfs2_dio_end_io */
2346                 ocfs2_iocb_set_rw_locked(iocb, rw_level);
2347         }
2348
2349         /*
2350          * We're fine letting folks race truncates and extending
2351          * writes with read across the cluster, just like they can
2352          * locally. Hence no rw_lock during read.
2353          *
2354          * Take and drop the meta data lock to update inode fields
2355          * like i_size. This allows the checks down below
2356          * generic_file_aio_read() a chance of actually working.
2357          */
2358         ret = ocfs2_inode_lock_atime(inode, filp->f_vfsmnt, &lock_level);
2359         if (ret < 0) {
2360                 mlog_errno(ret);
2361                 goto bail;
2362         }
2363         ocfs2_inode_unlock(inode, lock_level);
2364
2365         ret = generic_file_aio_read(iocb, iov, nr_segs, iocb->ki_pos);
2366         if (ret == -EINVAL)
2367                 mlog(0, "generic_file_aio_read returned -EINVAL\n");
2368
2369         /* buffered aio wouldn't have proper lock coverage today */
2370         BUG_ON(ret == -EIOCBQUEUED && !(filp->f_flags & O_DIRECT));
2371
2372         /* see ocfs2_file_aio_write */
2373         if (ret == -EIOCBQUEUED || !ocfs2_iocb_is_rw_locked(iocb)) {
2374                 rw_level = -1;
2375                 have_alloc_sem = 0;
2376         }
2377
2378 bail:
2379         if (have_alloc_sem)
2380                 up_read(&inode->i_alloc_sem);
2381         if (rw_level != -1)
2382                 ocfs2_rw_unlock(inode, rw_level);
2383         mlog_exit(ret);
2384
2385         return ret;
2386 }
2387
2388 const struct inode_operations ocfs2_file_iops = {
2389         .setattr        = ocfs2_setattr,
2390         .getattr        = ocfs2_getattr,
2391         .permission     = ocfs2_permission,
2392         .setxattr       = generic_setxattr,
2393         .getxattr       = generic_getxattr,
2394         .listxattr      = ocfs2_listxattr,
2395         .removexattr    = generic_removexattr,
2396         .fallocate      = ocfs2_fallocate,
2397         .fiemap         = ocfs2_fiemap,
2398 };
2399
2400 const struct inode_operations ocfs2_special_file_iops = {
2401         .setattr        = ocfs2_setattr,
2402         .getattr        = ocfs2_getattr,
2403         .permission     = ocfs2_permission,
2404 };
2405
2406 /*
2407  * Other than ->lock, keep ocfs2_fops and ocfs2_dops in sync with
2408  * ocfs2_fops_no_plocks and ocfs2_dops_no_plocks!
2409  */
2410 const struct file_operations ocfs2_fops = {
2411         .llseek         = generic_file_llseek,
2412         .read           = do_sync_read,
2413         .write          = do_sync_write,
2414         .mmap           = ocfs2_mmap,
2415         .fsync          = ocfs2_sync_file,
2416         .release        = ocfs2_file_release,
2417         .open           = ocfs2_file_open,
2418         .aio_read       = ocfs2_file_aio_read,
2419         .aio_write      = ocfs2_file_aio_write,
2420         .unlocked_ioctl = ocfs2_ioctl,
2421 #ifdef CONFIG_COMPAT
2422         .compat_ioctl   = ocfs2_compat_ioctl,
2423 #endif
2424         .lock           = ocfs2_lock,
2425         .flock          = ocfs2_flock,
2426         .splice_read    = ocfs2_file_splice_read,
2427         .splice_write   = ocfs2_file_splice_write,
2428 };
2429
2430 const struct file_operations ocfs2_dops = {
2431         .llseek         = generic_file_llseek,
2432         .read           = generic_read_dir,
2433         .readdir        = ocfs2_readdir,
2434         .fsync          = ocfs2_sync_file,
2435         .release        = ocfs2_dir_release,
2436         .open           = ocfs2_dir_open,
2437         .unlocked_ioctl = ocfs2_ioctl,
2438 #ifdef CONFIG_COMPAT
2439         .compat_ioctl   = ocfs2_compat_ioctl,
2440 #endif
2441         .lock           = ocfs2_lock,
2442         .flock          = ocfs2_flock,
2443 };
2444
2445 /*
2446  * POSIX-lockless variants of our file_operations.
2447  *
2448  * These will be used if the underlying cluster stack does not support
2449  * posix file locking, if the user passes the "localflocks" mount
2450  * option, or if we have a local-only fs.
2451  *
2452  * ocfs2_flock is in here because all stacks handle UNIX file locks,
2453  * so we still want it in the case of no stack support for
2454  * plocks. Internally, it will do the right thing when asked to ignore
2455  * the cluster.
2456  */
2457 const struct file_operations ocfs2_fops_no_plocks = {
2458         .llseek         = generic_file_llseek,
2459         .read           = do_sync_read,
2460         .write          = do_sync_write,
2461         .mmap           = ocfs2_mmap,
2462         .fsync          = ocfs2_sync_file,
2463         .release        = ocfs2_file_release,
2464         .open           = ocfs2_file_open,
2465         .aio_read       = ocfs2_file_aio_read,
2466         .aio_write      = ocfs2_file_aio_write,
2467         .unlocked_ioctl = ocfs2_ioctl,
2468 #ifdef CONFIG_COMPAT
2469         .compat_ioctl   = ocfs2_compat_ioctl,
2470 #endif
2471         .flock          = ocfs2_flock,
2472         .splice_read    = ocfs2_file_splice_read,
2473         .splice_write   = ocfs2_file_splice_write,
2474 };
2475
2476 const struct file_operations ocfs2_dops_no_plocks = {
2477         .llseek         = generic_file_llseek,
2478         .read           = generic_read_dir,
2479         .readdir        = ocfs2_readdir,
2480         .fsync          = ocfs2_sync_file,
2481         .release        = ocfs2_dir_release,
2482         .open           = ocfs2_dir_open,
2483         .unlocked_ioctl = ocfs2_ioctl,
2484 #ifdef CONFIG_COMPAT
2485         .compat_ioctl   = ocfs2_compat_ioctl,
2486 #endif
2487         .flock          = ocfs2_flock,
2488 };