fs/ocfs2/file.c

   1 /* -*- mode: c; c-basic-offset: 8; -*-
   2  * vim: noexpandtab sw=8 ts=8 sts=0:
   3  *
   4  * file.c
   5  *
   6  * File open, close, extend, truncate
   7  *
   8  * Copyright (C) 2002, 2004 Oracle.  All rights reserved.
   9  *
  10  * This program is free software; you can redistribute it and/or
  11  * modify it under the terms of the GNU General Public
  12  * License as published by the Free Software Foundation; either
  13  * version 2 of the License, or (at your option) any later version.
  14  *
  15  * This program is distributed in the hope that it will be useful,
  16  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  17  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  18  * General Public License for more details.
  19  *
  20  * You should have received a copy of the GNU General Public
  21  * License along with this program; if not, write to the
  22  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  23  * Boston, MA 021110-1307, USA.
  24  */
  25
  26 #include <linux/capability.h>
  27 #include <linux/fs.h>
  28 #include <linux/types.h>
  29 #include <linux/slab.h>
  30 #include <linux/highmem.h>
  31 #include <linux/pagemap.h>
  32 #include <linux/uio.h>
  33 #include <linux/sched.h>
  34 #include <linux/splice.h>
  35 #include <linux/mount.h>
  36 #include <linux/writeback.h>
  37 #include <linux/falloc.h>
  38 #include <linux/quotaops.h>
  39
  40 #define MLOG_MASK_PREFIX ML_INODE
  41 #include <cluster/masklog.h>
  42
  43 #include "ocfs2.h"
  44
  45 #include "alloc.h"
  46 #include "aops.h"
  47 #include "dir.h"
  48 #include "dlmglue.h"
  49 #include "extent_map.h"
  50 #include "file.h"
  51 #include "sysfile.h"
  52 #include "inode.h"
  53 #include "ioctl.h"
  54 #include "journal.h"
  55 #include "locks.h"
  56 #include "mmap.h"
  57 #include "suballoc.h"
  58 #include "super.h"
  59 #include "xattr.h"
  60 #include "acl.h"
  61 #include "quota.h"
  62 #include "refcounttree.h"
  63
  64 #include "buffer_head_io.h"
  65
  66 static int ocfs2_sync_inode(struct inode *inode)
  67 {
  68         filemap_fdatawrite(inode->i_mapping);
  69         return sync_mapping_buffers(inode->i_mapping);
  70 }
  71
  72 static int ocfs2_init_file_private(struct inode *inode, struct file *file)
  73 {
  74         struct ocfs2_file_private *fp;
  75
  76         fp = kzalloc(sizeof(struct ocfs2_file_private), GFP_KERNEL);
  77         if (!fp)
  78                 return -ENOMEM;
  79
  80         fp->fp_file = file;
  81         mutex_init(&fp->fp_mutex);
  82         ocfs2_file_lock_res_init(&fp->fp_flock, fp);
  83         file->private_data = fp;
  84
  85         return 0;
  86 }
  87
  88 static void ocfs2_free_file_private(struct inode *inode, struct file *file)
  89 {
  90         struct ocfs2_file_private *fp = file->private_data;
  91         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
  92
  93         if (fp) {
  94                 ocfs2_simple_drop_lockres(osb, &fp->fp_flock);
  95                 ocfs2_lock_res_free(&fp->fp_flock);
  96                 kfree(fp);
  97                 file->private_data = NULL;
  98         }
  99 }
 100
 101 static int ocfs2_file_open(struct inode *inode, struct file *file)
 102 {
 103         int status;
 104         int mode = file->f_flags;
 105         struct ocfs2_inode_info *oi = OCFS2_I(inode);
 106
 107         mlog_entry("(0x%p, 0x%p, '%.*s')\n", inode, file,
 108                    file->f_path.dentry->d_name.len, file->f_path.dentry->d_name.name);
 109
 110         spin_lock(&oi->ip_lock);
 111
 112         /* Check that the inode hasn't been wiped from disk by another
 113          * node. If it hasn't then we're safe as long as we hold the
 114          * spin lock until our increment of open count. */
 115         if (OCFS2_I(inode)->ip_flags & OCFS2_INODE_DELETED) {
 116                 spin_unlock(&oi->ip_lock);
 117
 118                 status = -ENOENT;
 119                 goto leave;
 120         }
 121
 122         if (mode & O_DIRECT)
 123                 oi->ip_flags |= OCFS2_INODE_OPEN_DIRECT;
 124
 125         oi->ip_open_count++;
 126         spin_unlock(&oi->ip_lock);
 127
 128         status = ocfs2_init_file_private(inode, file);
 129         if (status) {
 130                 /*
 131                  * We want to set open count back if we're failing the
 132                  * open.
 133                  */
 134                 spin_lock(&oi->ip_lock);
 135                 oi->ip_open_count--;
 136                 spin_unlock(&oi->ip_lock);
 137         }
 138
 139 leave:
 140         mlog_exit(status);
 141         return status;
 142 }
 143
 144 static int ocfs2_file_release(struct inode *inode, struct file *file)
 145 {
 146         struct ocfs2_inode_info *oi = OCFS2_I(inode);
 147         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 148
 149         mlog_entry("(0x%p, 0x%p, '%.*s')\n", inode, file,
 150                        file->f_path.dentry->d_name.len,
 151                        file->f_path.dentry->d_name.name);
 152
 153         spin_lock(&oi->ip_lock);
 154         if (!--oi->ip_open_count)
 155                 oi->ip_flags &= ~OCFS2_INODE_OPEN_DIRECT;
 156         spin_unlock(&oi->ip_lock);
 157
 158 #if 0
 159         /*
 160          * Disable this for now. Keeping the reservation around a bit
 161          * longer gives an improvement for workloads which rapidly do
 162          * open()/write()/close() against a file.
 163          */
 164         if ((file->f_mode & FMODE_WRITE) &&
 165             (atomic_read(&inode->i_writecount) == 1)) {
 166                 down_write(&oi->ip_alloc_sem);
 167                 ocfs2_resv_discard(&osb->osb_la_resmap,
 168                                    &oi->ip_la_data_resv);
 169                 up_write(&oi->ip_alloc_sem);
 170         }
 171 #endif
 172
 173         ocfs2_free_file_private(inode, file);
 174
 175         mlog_exit(0);
 176
 177         return 0;
 178 }
 179
 180 static int ocfs2_dir_open(struct inode *inode, struct file *file)
 181 {
 182         return ocfs2_init_file_private(inode, file);
 183 }
 184
 185 static int ocfs2_dir_release(struct inode *inode, struct file *file)
 186 {
 187         ocfs2_free_file_private(inode, file);
 188         return 0;
 189 }
 190
 191 static int ocfs2_sync_file(struct file *file,
 192                            struct dentry *dentry,
 193                            int datasync)
 194 {
 195         int err = 0;
 196         journal_t *journal;
 197         struct inode *inode = dentry->d_inode;
 198         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 199
 200         mlog_entry("(0x%p, 0x%p, %d, '%.*s')\n", file, dentry, datasync,
 201                    dentry->d_name.len, dentry->d_name.name);
 202
 203         err = ocfs2_sync_inode(dentry->d_inode);
 204         if (err)
 205                 goto bail;
 206
 207         if (datasync && !(inode->i_state & I_DIRTY_DATASYNC))
 208                 goto bail;
 209
 210         journal = osb->journal->j_journal;
 211         err = jbd2_journal_force_commit(journal);
 212
 213 bail:
 214         mlog_exit(err);
 215
 216         return (err < 0) ? -EIO : 0;
 217 }
 218
 219 int ocfs2_should_update_atime(struct inode *inode,
 220                               struct vfsmount *vfsmnt)
 221 {
 222         struct timespec now;
 223         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 224
 225         if (ocfs2_is_hard_readonly(osb) || ocfs2_is_soft_readonly(osb))
 226                 return 0;
 227
 228         if ((inode->i_flags & S_NOATIME) ||
 229             ((inode->i_sb->s_flags & MS_NODIRATIME) && S_ISDIR(inode->i_mode)))
 230                 return 0;
 231
 232         /*
 233          * We can be called with no vfsmnt structure - NFSD will
 234          * sometimes do this.
 235          *
 236          * Note that our action here is different than touch_atime() -
 237          * if we can't tell whether this is a noatime mount, then we
 238          * don't know whether to trust the value of s_atime_quantum.
 239          */
 240         if (vfsmnt == NULL)
 241                 return 0;
 242
 243         if ((vfsmnt->mnt_flags & MNT_NOATIME) ||
 244             ((vfsmnt->mnt_flags & MNT_NODIRATIME) && S_ISDIR(inode->i_mode)))
 245                 return 0;
 246
 247         if (vfsmnt->mnt_flags & MNT_RELATIME) {
 248                 if ((timespec_compare(&inode->i_atime, &inode->i_mtime) <= 0) ||
 249                     (timespec_compare(&inode->i_atime, &inode->i_ctime) <= 0))
 250                         return 1;
 251
 252                 return 0;
 253         }
 254
 255         now = CURRENT_TIME;
 256         if ((now.tv_sec - inode->i_atime.tv_sec <= osb->s_atime_quantum))
 257                 return 0;
 258         else
 259                 return 1;
 260 }
 261
 262 int ocfs2_update_inode_atime(struct inode *inode,
 263                              struct buffer_head *bh)
 264 {
 265         int ret;
 266         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 267         handle_t *handle;
 268         struct ocfs2_dinode *di = (struct ocfs2_dinode *) bh->b_data;
 269
 270         mlog_entry_void();
 271
 272         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
 273         if (IS_ERR(handle)) {
 274                 ret = PTR_ERR(handle);
 275                 mlog_errno(ret);
 276                 goto out;
 277         }
 278
 279         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), bh,
 280                                       OCFS2_JOURNAL_ACCESS_WRITE);
 281         if (ret) {
 282                 mlog_errno(ret);
 283                 goto out_commit;
 284         }
 285
 286         /*
 287          * Don't use ocfs2_mark_inode_dirty() here as we don't always
 288          * have i_mutex to guard against concurrent changes to other
 289          * inode fields.
 290          */
 291         inode->i_atime = CURRENT_TIME;
 292         di->i_atime = cpu_to_le64(inode->i_atime.tv_sec);
 293         di->i_atime_nsec = cpu_to_le32(inode->i_atime.tv_nsec);
 294
 295         ret = ocfs2_journal_dirty(handle, bh);
 296         if (ret < 0)
 297                 mlog_errno(ret);
 298
 299 out_commit:
 300         ocfs2_commit_trans(OCFS2_SB(inode->i_sb), handle);
 301 out:
 302         mlog_exit(ret);
 303         return ret;
 304 }
 305
 306 static int ocfs2_set_inode_size(handle_t *handle,
 307                                 struct inode *inode,
 308                                 struct buffer_head *fe_bh,
 309                                 u64 new_i_size)
 310 {
 311         int status;
 312
 313         mlog_entry_void();
 314         i_size_write(inode, new_i_size);
 315         inode->i_blocks = ocfs2_inode_sector_count(inode);
 316         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
 317
 318         status = ocfs2_mark_inode_dirty(handle, inode, fe_bh);
 319         if (status < 0) {
 320                 mlog_errno(status);
 321                 goto bail;
 322         }
 323
 324 bail:
 325         mlog_exit(status);
 326         return status;
 327 }
 328
 329 int ocfs2_simple_size_update(struct inode *inode,
 330                              struct buffer_head *di_bh,
 331                              u64 new_i_size)
 332 {
 333         int ret;
 334         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 335         handle_t *handle = NULL;
 336
 337         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
 338         if (IS_ERR(handle)) {
 339                 ret = PTR_ERR(handle);
 340                 mlog_errno(ret);
 341                 goto out;
 342         }
 343
 344         ret = ocfs2_set_inode_size(handle, inode, di_bh,
 345                                    new_i_size);
 346         if (ret < 0)
 347                 mlog_errno(ret);
 348
 349         ocfs2_commit_trans(osb, handle);
 350 out:
 351         return ret;
 352 }
 353
 354 static int ocfs2_cow_file_pos(struct inode *inode,
 355                               struct buffer_head *fe_bh,
 356                               u64 offset)
 357 {
 358         int status;
 359         u32 phys, cpos = offset >> OCFS2_SB(inode->i_sb)->s_clustersize_bits;
 360         unsigned int num_clusters = 0;
 361         unsigned int ext_flags = 0;
 362
 363         /*
 364          * If the new offset is aligned to the range of the cluster, there is
 365          * no space for ocfs2_zero_range_for_truncate to fill, so no need to
 366          * CoW either.
 367          */
 368         if ((offset & (OCFS2_SB(inode->i_sb)->s_clustersize - 1)) == 0)
 369                 return 0;
 370
 371         status = ocfs2_get_clusters(inode, cpos, &phys,
 372                                     &num_clusters, &ext_flags);
 373         if (status) {
 374                 mlog_errno(status);
 375                 goto out;
 376         }
 377
 378         if (!(ext_flags & OCFS2_EXT_REFCOUNTED))
 379                 goto out;
 380
 381         return ocfs2_refcount_cow(inode, fe_bh, cpos, 1, cpos+1);
 382
 383 out:
 384         return status;
 385 }
 386
 387 static int ocfs2_orphan_for_truncate(struct ocfs2_super *osb,
 388                                      struct inode *inode,
 389                                      struct buffer_head *fe_bh,
 390                                      u64 new_i_size)
 391 {
 392         int status;
 393         handle_t *handle;
 394         struct ocfs2_dinode *di;
 395         u64 cluster_bytes;
 396
 397         mlog_entry_void();
 398
 399         /*
 400          * We need to CoW the cluster contains the offset if it is reflinked
 401          * since we will call ocfs2_zero_range_for_truncate later which will
 402          * write "0" from offset to the end of the cluster.
 403          */
 404         status = ocfs2_cow_file_pos(inode, fe_bh, new_i_size);
 405         if (status) {
 406                 mlog_errno(status);
 407                 return status;
 408         }
 409
 410         /* TODO: This needs to actually orphan the inode in this
 411          * transaction. */
 412
 413         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
 414         if (IS_ERR(handle)) {
 415                 status = PTR_ERR(handle);
 416                 mlog_errno(status);
 417                 goto out;
 418         }
 419
 420         status = ocfs2_journal_access_di(handle, INODE_CACHE(inode), fe_bh,
 421                                          OCFS2_JOURNAL_ACCESS_WRITE);
 422         if (status < 0) {
 423                 mlog_errno(status);
 424                 goto out_commit;
 425         }
 426
 427         /*
 428          * Do this before setting i_size.
 429          */
 430         cluster_bytes = ocfs2_align_bytes_to_clusters(inode->i_sb, new_i_size);
 431         status = ocfs2_zero_range_for_truncate(inode, handle, new_i_size,
 432                                                cluster_bytes);
 433         if (status) {
 434                 mlog_errno(status);
 435                 goto out_commit;
 436         }
 437
 438         i_size_write(inode, new_i_size);
 439         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
 440
 441         di = (struct ocfs2_dinode *) fe_bh->b_data;
 442         di->i_size = cpu_to_le64(new_i_size);
 443         di->i_ctime = di->i_mtime = cpu_to_le64(inode->i_ctime.tv_sec);
 444         di->i_ctime_nsec = di->i_mtime_nsec = cpu_to_le32(inode->i_ctime.tv_nsec);
 445
 446         status = ocfs2_journal_dirty(handle, fe_bh);
 447         if (status < 0)
 448                 mlog_errno(status);
 449
 450 out_commit:
 451         ocfs2_commit_trans(osb, handle);
 452 out:
 453
 454         mlog_exit(status);
 455         return status;
 456 }
 457
 458 static int ocfs2_truncate_file(struct inode *inode,
 459                                struct buffer_head *di_bh,
 460                                u64 new_i_size)
 461 {
 462         int status = 0;
 463         struct ocfs2_dinode *fe = NULL;
 464         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 465         struct ocfs2_truncate_context *tc = NULL;
 466
 467         mlog_entry("(inode = %llu, new_i_size = %llu\n",
 468                    (unsigned long long)OCFS2_I(inode)->ip_blkno,
 469                    (unsigned long long)new_i_size);
 470
 471         /* We trust di_bh because it comes from ocfs2_inode_lock(), which
 472          * already validated it */
 473         fe = (struct ocfs2_dinode *) di_bh->b_data;
 474
 475         mlog_bug_on_msg(le64_to_cpu(fe->i_size) != i_size_read(inode),
 476                         "Inode %llu, inode i_size = %lld != di "
 477                         "i_size = %llu, i_flags = 0x%x\n",
 478                         (unsigned long long)OCFS2_I(inode)->ip_blkno,
 479                         i_size_read(inode),
 480                         (unsigned long long)le64_to_cpu(fe->i_size),
 481                         le32_to_cpu(fe->i_flags));
 482
 483         if (new_i_size > le64_to_cpu(fe->i_size)) {
 484                 mlog(0, "asked to truncate file with size (%llu) to size (%llu)!\n",
 485                      (unsigned long long)le64_to_cpu(fe->i_size),
 486                      (unsigned long long)new_i_size);
 487                 status = -EINVAL;
 488                 mlog_errno(status);
 489                 goto bail;
 490         }
 491
 492         mlog(0, "inode %llu, i_size = %llu, new_i_size = %llu\n",
 493              (unsigned long long)le64_to_cpu(fe->i_blkno),
 494              (unsigned long long)le64_to_cpu(fe->i_size),
 495              (unsigned long long)new_i_size);
 496
 497         /* lets handle the simple truncate cases before doing any more
 498          * cluster locking. */
 499         if (new_i_size == le64_to_cpu(fe->i_size))
 500                 goto bail;
 501
 502         down_write(&OCFS2_I(inode)->ip_alloc_sem);
 503
 504         ocfs2_resv_discard(&osb->osb_la_resmap,
 505                            &OCFS2_I(inode)->ip_la_data_resv);
 506
 507         /*
 508          * The inode lock forced other nodes to sync and drop their
 509          * pages, which (correctly) happens even if we have a truncate
 510          * without allocation change - ocfs2 cluster sizes can be much
 511          * greater than page size, so we have to truncate them
 512          * anyway.
 513          */
 514         unmap_mapping_range(inode->i_mapping, new_i_size + PAGE_SIZE - 1, 0, 1);
 515         truncate_inode_pages(inode->i_mapping, new_i_size);
 516
 517         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
 518                 status = ocfs2_truncate_inline(inode, di_bh, new_i_size,
 519                                                i_size_read(inode), 1);
 520                 if (status)
 521                         mlog_errno(status);
 522
 523                 goto bail_unlock_sem;
 524         }
 525
 526         /* alright, we're going to need to do a full blown alloc size
 527          * change. Orphan the inode so that recovery can complete the
 528          * truncate if necessary. This does the task of marking
 529          * i_size. */
 530         status = ocfs2_orphan_for_truncate(osb, inode, di_bh, new_i_size);
 531         if (status < 0) {
 532                 mlog_errno(status);
 533                 goto bail_unlock_sem;
 534         }
 535
 536         status = ocfs2_prepare_truncate(osb, inode, di_bh, &tc);
 537         if (status < 0) {
 538                 mlog_errno(status);
 539                 goto bail_unlock_sem;
 540         }
 541
 542         status = ocfs2_commit_truncate(osb, inode, di_bh, tc);
 543         if (status < 0) {
 544                 mlog_errno(status);
 545                 goto bail_unlock_sem;
 546         }
 547
 548         /* TODO: orphan dir cleanup here. */
 549 bail_unlock_sem:
 550         up_write(&OCFS2_I(inode)->ip_alloc_sem);
 551
 552 bail:
 553         if (!status && OCFS2_I(inode)->ip_clusters == 0)
 554                 status = ocfs2_try_remove_refcount_tree(inode, di_bh);
 555
 556         mlog_exit(status);
 557         return status;
 558 }
 559
 560 /*
 561  * extend file allocation only here.
 562  * we'll update all the disk stuff, and oip->alloc_size
 563  *
 564  * expect stuff to be locked, a transaction started and enough data /
 565  * metadata reservations in the contexts.
 566  *
 567  * Will return -EAGAIN, and a reason if a restart is needed.
 568  * If passed in, *reason will always be set, even in error.
 569  */
 570 int ocfs2_add_inode_data(struct ocfs2_super *osb,
 571                          struct inode *inode,
 572                          u32 *logical_offset,
 573                          u32 clusters_to_add,
 574                          int mark_unwritten,
 575                          struct buffer_head *fe_bh,
 576                          handle_t *handle,
 577                          struct ocfs2_alloc_context *data_ac,
 578                          struct ocfs2_alloc_context *meta_ac,
 579                          enum ocfs2_alloc_restarted *reason_ret)
 580 {
 581         int ret;
 582         struct ocfs2_extent_tree et;
 583
 584         ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), fe_bh);
 585         ret = ocfs2_add_clusters_in_btree(handle, &et, logical_offset,
 586                                           clusters_to_add, mark_unwritten,
 587                                           data_ac, meta_ac, reason_ret);
 588
 589         return ret;
 590 }
 591
 592 static int __ocfs2_extend_allocation(struct inode *inode, u32 logical_start,
 593                                      u32 clusters_to_add, int mark_unwritten)
 594 {
 595         int status = 0;
 596         int restart_func = 0;
 597         int credits;
 598         u32 prev_clusters;
 599         struct buffer_head *bh = NULL;
 600         struct ocfs2_dinode *fe = NULL;
 601         handle_t *handle = NULL;
 602         struct ocfs2_alloc_context *data_ac = NULL;
 603         struct ocfs2_alloc_context *meta_ac = NULL;
 604         enum ocfs2_alloc_restarted why;
 605         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 606         struct ocfs2_extent_tree et;
 607         int did_quota = 0;
 608
 609         mlog_entry("(clusters_to_add = %u)\n", clusters_to_add);
 610
 611         /*
 612          * This function only exists for file systems which don't
 613          * support holes.
 614          */
 615         BUG_ON(mark_unwritten && !ocfs2_sparse_alloc(osb));
 616
 617         status = ocfs2_read_inode_block(inode, &bh);
 618         if (status < 0) {
 619                 mlog_errno(status);
 620                 goto leave;
 621         }
 622         fe = (struct ocfs2_dinode *) bh->b_data;
 623
 624 restart_all:
 625         BUG_ON(le32_to_cpu(fe->i_clusters) != OCFS2_I(inode)->ip_clusters);
 626
 627         mlog(0, "extend inode %llu, i_size = %lld, di->i_clusters = %u, "
 628              "clusters_to_add = %u\n",
 629              (unsigned long long)OCFS2_I(inode)->ip_blkno,
 630              (long long)i_size_read(inode), le32_to_cpu(fe->i_clusters),
 631              clusters_to_add);
 632         ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), bh);
 633         status = ocfs2_lock_allocators(inode, &et, clusters_to_add, 0,
 634                                        &data_ac, &meta_ac);
 635         if (status) {
 636                 mlog_errno(status);
 637                 goto leave;
 638         }
 639
 640         credits = ocfs2_calc_extend_credits(osb->sb, &fe->id2.i_list,
 641                                             clusters_to_add);
 642         handle = ocfs2_start_trans(osb, credits);
 643         if (IS_ERR(handle)) {
 644                 status = PTR_ERR(handle);
 645                 handle = NULL;
 646                 mlog_errno(status);
 647                 goto leave;
 648         }
 649
 650 restarted_transaction:
 651         if (vfs_dq_alloc_space_nodirty(inode, ocfs2_clusters_to_bytes(osb->sb,
 652             clusters_to_add))) {
 653                 status = -EDQUOT;
 654                 goto leave;
 655         }
 656         did_quota = 1;
 657
 658         /* reserve a write to the file entry early on - that we if we
 659          * run out of credits in the allocation path, we can still
 660          * update i_size. */
 661         status = ocfs2_journal_access_di(handle, INODE_CACHE(inode), bh,
 662                                          OCFS2_JOURNAL_ACCESS_WRITE);
 663         if (status < 0) {
 664                 mlog_errno(status);
 665                 goto leave;
 666         }
 667
 668         prev_clusters = OCFS2_I(inode)->ip_clusters;
 669
 670         status = ocfs2_add_inode_data(osb,
 671                                       inode,
 672                                       &logical_start,
 673                                       clusters_to_add,
 674                                       mark_unwritten,
 675                                       bh,
 676                                       handle,
 677                                       data_ac,
 678                                       meta_ac,
 679                                       &why);
 680         if ((status < 0) && (status != -EAGAIN)) {
 681                 if (status != -ENOSPC)
 682                         mlog_errno(status);
 683                 goto leave;
 684         }
 685
 686         status = ocfs2_journal_dirty(handle, bh);
 687         if (status < 0) {
 688                 mlog_errno(status);
 689                 goto leave;
 690         }
 691
 692         spin_lock(&OCFS2_I(inode)->ip_lock);
 693         clusters_to_add -= (OCFS2_I(inode)->ip_clusters - prev_clusters);
 694         spin_unlock(&OCFS2_I(inode)->ip_lock);
 695         /* Release unused quota reservation */
 696         vfs_dq_free_space(inode,
 697                         ocfs2_clusters_to_bytes(osb->sb, clusters_to_add));
 698         did_quota = 0;
 699
 700         if (why != RESTART_NONE && clusters_to_add) {
 701                 if (why == RESTART_META) {
 702                         mlog(0, "restarting function.\n");
 703                         restart_func = 1;
 704                 } else {
 705                         BUG_ON(why != RESTART_TRANS);
 706
 707                         mlog(0, "restarting transaction.\n");
 708                         /* TODO: This can be more intelligent. */
 709                         credits = ocfs2_calc_extend_credits(osb->sb,
 710                                                             &fe->id2.i_list,
 711                                                             clusters_to_add);
 712                         status = ocfs2_extend_trans(handle, credits);
 713                         if (status < 0) {
 714                                 /* handle still has to be committed at
 715                                  * this point. */
 716                                 status = -ENOMEM;
 717                                 mlog_errno(status);
 718                                 goto leave;
 719                         }
 720                         goto restarted_transaction;
 721                 }
 722         }
 723
 724         mlog(0, "fe: i_clusters = %u, i_size=%llu\n",
 725              le32_to_cpu(fe->i_clusters),
 726              (unsigned long long)le64_to_cpu(fe->i_size));
 727         mlog(0, "inode: ip_clusters=%u, i_size=%lld\n",
 728              OCFS2_I(inode)->ip_clusters, (long long)i_size_read(inode));
 729
 730 leave:
 731         if (status < 0 && did_quota)
 732                 vfs_dq_free_space(inode,
 733                         ocfs2_clusters_to_bytes(osb->sb, clusters_to_add));
 734         if (handle) {
 735                 ocfs2_commit_trans(osb, handle);
 736                 handle = NULL;
 737         }
 738         if (data_ac) {
 739                 ocfs2_free_alloc_context(data_ac);
 740                 data_ac = NULL;
 741         }
 742         if (meta_ac) {
 743                 ocfs2_free_alloc_context(meta_ac);
 744                 meta_ac = NULL;
 745         }
 746         if ((!status) && restart_func) {
 747                 restart_func = 0;
 748                 goto restart_all;
 749         }
 750         brelse(bh);
 751         bh = NULL;
 752
 753         mlog_exit(status);
 754         return status;
 755 }
 756
 757 /* Some parts of this taken from generic_cont_expand, which turned out
 758  * to be too fragile to do exactly what we need without us having to
 759  * worry about recursive locking in ->write_begin() and ->write_end(). */
 760 static int ocfs2_write_zero_page(struct inode *inode,
 761                                  u64 size)
 762 {
 763         struct address_space *mapping = inode->i_mapping;
 764         struct page *page;
 765         unsigned long index;
 766         unsigned int offset;
 767         handle_t *handle = NULL;
 768         int ret;
 769
 770         offset = (size & (PAGE_CACHE_SIZE-1)); /* Within page */
 771         /* ugh.  in prepare/commit_write, if from==to==start of block, we
 772         ** skip the prepare.  make sure we never send an offset for the start
 773         ** of a block
 774         */
 775         if ((offset & (inode->i_sb->s_blocksize - 1)) == 0) {
 776                 offset++;
 777         }
 778         index = size >> PAGE_CACHE_SHIFT;
 779
 780         page = grab_cache_page(mapping, index);
 781         if (!page) {
 782                 ret = -ENOMEM;
 783                 mlog_errno(ret);
 784                 goto out;
 785         }
 786
 787         ret = ocfs2_prepare_write_nolock(inode, page, offset, offset);
 788         if (ret < 0) {
 789                 mlog_errno(ret);
 790                 goto out_unlock;
 791         }
 792
 793         if (ocfs2_should_order_data(inode)) {
 794                 handle = ocfs2_start_walk_page_trans(inode, page, offset,
 795                                                      offset);
 796                 if (IS_ERR(handle)) {
 797                         ret = PTR_ERR(handle);
 798                         handle = NULL;
 799                         goto out_unlock;
 800                 }
 801         }
 802
 803         /* must not update i_size! */
 804         ret = block_commit_write(page, offset, offset);
 805         if (ret < 0)
 806                 mlog_errno(ret);
 807         else
 808                 ret = 0;
 809
 810         if (handle)
 811                 ocfs2_commit_trans(OCFS2_SB(inode->i_sb), handle);
 812 out_unlock:
 813         unlock_page(page);
 814         page_cache_release(page);
 815 out:
 816         return ret;
 817 }
 818
 819 static int ocfs2_zero_extend(struct inode *inode,
 820                              u64 zero_to_size)
 821 {
 822         int ret = 0;
 823         u64 start_off;
 824         struct super_block *sb = inode->i_sb;
 825
 826         start_off = ocfs2_align_bytes_to_blocks(sb, i_size_read(inode));
 827         while (start_off < zero_to_size) {
 828                 ret = ocfs2_write_zero_page(inode, start_off);
 829                 if (ret < 0) {
 830                         mlog_errno(ret);
 831                         goto out;
 832                 }
 833
 834                 start_off += sb->s_blocksize;
 835
 836                 /*
 837                  * Very large extends have the potential to lock up
 838                  * the cpu for extended periods of time.
 839                  */
 840                 cond_resched();
 841         }
 842
 843 out:
 844         return ret;
 845 }
 846
 847 int ocfs2_extend_no_holes(struct inode *inode, u64 new_i_size, u64 zero_to)
 848 {
 849         int ret;
 850         u32 clusters_to_add;
 851         struct ocfs2_inode_info *oi = OCFS2_I(inode);
 852
 853         clusters_to_add = ocfs2_clusters_for_bytes(inode->i_sb, new_i_size);
 854         if (clusters_to_add < oi->ip_clusters)
 855                 clusters_to_add = 0;
 856         else
 857                 clusters_to_add -= oi->ip_clusters;
 858
 859         if (clusters_to_add) {
 860                 ret = __ocfs2_extend_allocation(inode, oi->ip_clusters,
 861                                                 clusters_to_add, 0);
 862                 if (ret) {
 863                         mlog_errno(ret);
 864                         goto out;
 865                 }
 866         }
 867
 868         /*
 869          * Call this even if we don't add any clusters to the tree. We
 870          * still need to zero the area between the old i_size and the
 871          * new i_size.
 872          */
 873         ret = ocfs2_zero_extend(inode, zero_to);
 874         if (ret < 0)
 875                 mlog_errno(ret);
 876
 877 out:
 878         return ret;
 879 }
 880
 881 static int ocfs2_extend_file(struct inode *inode,
 882                              struct buffer_head *di_bh,
 883                              u64 new_i_size)
 884 {
 885         int ret = 0;
 886         struct ocfs2_inode_info *oi = OCFS2_I(inode);
 887
 888         BUG_ON(!di_bh);
 889
 890         /* setattr sometimes calls us like this. */
 891         if (new_i_size == 0)
 892                 goto out;
 893
 894         if (i_size_read(inode) == new_i_size)
 895                 goto out;
 896         BUG_ON(new_i_size < i_size_read(inode));
 897
 898         /*
 899          * Fall through for converting inline data, even if the fs
 900          * supports sparse files.
 901          *
 902          * The check for inline data here is legal - nobody can add
 903          * the feature since we have i_mutex. We must check it again
 904          * after acquiring ip_alloc_sem though, as paths like mmap
 905          * might have raced us to converting the inode to extents.
 906          */
 907         if (!(oi->ip_dyn_features & OCFS2_INLINE_DATA_FL)
 908             && ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)))
 909                 goto out_update_size;
 910
 911         /*
 912          * The alloc sem blocks people in read/write from reading our
 913          * allocation until we're done changing it. We depend on
 914          * i_mutex to block other extend/truncate calls while we're
 915          * here.
 916          */
 917         down_write(&oi->ip_alloc_sem);
 918
 919         if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
 920                 /*
 921                  * We can optimize small extends by keeping the inodes
 922                  * inline data.
 923                  */
 924                 if (ocfs2_size_fits_inline_data(di_bh, new_i_size)) {
 925                         up_write(&oi->ip_alloc_sem);
 926                         goto out_update_size;
 927                 }
 928
 929                 ret = ocfs2_convert_inline_data_to_extents(inode, di_bh);
 930                 if (ret) {
 931                         up_write(&oi->ip_alloc_sem);
 932
 933                         mlog_errno(ret);
 934                         goto out;
 935                 }
 936         }
 937
 938         if (!ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)))
 939                 ret = ocfs2_extend_no_holes(inode, new_i_size, new_i_size);
 940
 941         up_write(&oi->ip_alloc_sem);
 942
 943         if (ret < 0) {
 944                 mlog_errno(ret);
 945                 goto out;
 946         }
 947
 948 out_update_size:
 949         ret = ocfs2_simple_size_update(inode, di_bh, new_i_size);
 950         if (ret < 0)
 951                 mlog_errno(ret);
 952
 953 out:
 954         return ret;
 955 }
 956
 957 int ocfs2_setattr(struct dentry *dentry, struct iattr *attr)
 958 {
 959         int status = 0, size_change;
 960         struct inode *inode = dentry->d_inode;
 961         struct super_block *sb = inode->i_sb;
 962         struct ocfs2_super *osb = OCFS2_SB(sb);
 963         struct buffer_head *bh = NULL;
 964         handle_t *handle = NULL;
 965         int qtype;
 966         struct dquot *transfer_from[MAXQUOTAS] = { };
 967         struct dquot *transfer_to[MAXQUOTAS] = { };
 968
 969         mlog_entry("(0x%p, '%.*s')\n", dentry,
 970                    dentry->d_name.len, dentry->d_name.name);
 971
 972         /* ensuring we don't even attempt to truncate a symlink */
 973         if (S_ISLNK(inode->i_mode))
 974                 attr->ia_valid &= ~ATTR_SIZE;
 975
 976         if (attr->ia_valid & ATTR_MODE)
 977                 mlog(0, "mode change: %d\n", attr->ia_mode);
 978         if (attr->ia_valid & ATTR_UID)
 979                 mlog(0, "uid change: %d\n", attr->ia_uid);
 980         if (attr->ia_valid & ATTR_GID)
 981                 mlog(0, "gid change: %d\n", attr->ia_gid);
 982         if (attr->ia_valid & ATTR_SIZE)
 983                 mlog(0, "size change...\n");
 984         if (attr->ia_valid & (ATTR_ATIME | ATTR_MTIME | ATTR_CTIME))
 985                 mlog(0, "time change...\n");
 986
 987 #define OCFS2_VALID_ATTRS (ATTR_ATIME | ATTR_MTIME | ATTR_CTIME | ATTR_SIZE \
 988                            | ATTR_GID | ATTR_UID | ATTR_MODE)
 989         if (!(attr->ia_valid & OCFS2_VALID_ATTRS)) {
 990                 mlog(0, "can't handle attrs: 0x%x\n", attr->ia_valid);
 991                 return 0;
 992         }
 993
 994         status = inode_change_ok(inode, attr);
 995         if (status)
 996                 return status;
 997
 998         size_change = S_ISREG(inode->i_mode) && attr->ia_valid & ATTR_SIZE;
 999         if (size_change) {
1000                 status = ocfs2_rw_lock(inode, 1);
1001                 if (status < 0) {
1002                         mlog_errno(status);
1003                         goto bail;
1004                 }
1005         }
1006
1007         status = ocfs2_inode_lock(inode, &bh, 1);
1008         if (status < 0) {
1009                 if (status != -ENOENT)
1010                         mlog_errno(status);
1011                 goto bail_unlock_rw;
1012         }
1013
1014         if (size_change && attr->ia_size != i_size_read(inode)) {
1015                 if (attr->ia_size > sb->s_maxbytes) {
1016                         status = -EFBIG;
1017                         goto bail_unlock;
1018                 }
1019
1020                 if (i_size_read(inode) > attr->ia_size) {
1021                         if (ocfs2_should_order_data(inode)) {
1022                                 status = ocfs2_begin_ordered_truncate(inode,
1023                                                                       attr->ia_size);
1024                                 if (status)
1025                                         goto bail_unlock;
1026                         }
1027                         status = ocfs2_truncate_file(inode, bh, attr->ia_size);
1028                 } else
1029                         status = ocfs2_extend_file(inode, bh, attr->ia_size);
1030                 if (status < 0) {
1031                         if (status != -ENOSPC)
1032                                 mlog_errno(status);
1033                         status = -ENOSPC;
1034                         goto bail_unlock;
1035                 }
1036         }
1037
1038         if ((attr->ia_valid & ATTR_UID && attr->ia_uid != inode->i_uid) ||
1039             (attr->ia_valid & ATTR_GID && attr->ia_gid != inode->i_gid)) {
1040                 /*
1041                  * Gather pointers to quota structures so that allocation /
1042                  * freeing of quota structures happens here and not inside
1043                  * vfs_dq_transfer() where we have problems with lock ordering
1044                  */
1045                 if (attr->ia_valid & ATTR_UID && attr->ia_uid != inode->i_uid
1046                     && OCFS2_HAS_RO_COMPAT_FEATURE(sb,
1047                     OCFS2_FEATURE_RO_COMPAT_USRQUOTA)) {
1048                         transfer_to[USRQUOTA] = dqget(sb, attr->ia_uid,
1049                                                       USRQUOTA);
1050                         transfer_from[USRQUOTA] = dqget(sb, inode->i_uid,
1051                                                         USRQUOTA);
1052                         if (!transfer_to[USRQUOTA] || !transfer_from[USRQUOTA]) {
1053                                 status = -ESRCH;
1054                                 goto bail_unlock;
1055                         }
1056                 }
1057                 if (attr->ia_valid & ATTR_GID && attr->ia_gid != inode->i_gid
1058                     && OCFS2_HAS_RO_COMPAT_FEATURE(sb,
1059                     OCFS2_FEATURE_RO_COMPAT_GRPQUOTA)) {
1060                         transfer_to[GRPQUOTA] = dqget(sb, attr->ia_gid,
1061                                                       GRPQUOTA);
1062                         transfer_from[GRPQUOTA] = dqget(sb, inode->i_gid,
1063                                                         GRPQUOTA);
1064                         if (!transfer_to[GRPQUOTA] || !transfer_from[GRPQUOTA]) {
1065                                 status = -ESRCH;
1066                                 goto bail_unlock;
1067                         }
1068                 }
1069                 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS +
1070                                            2 * ocfs2_quota_trans_credits(sb));
1071                 if (IS_ERR(handle)) {
1072                         status = PTR_ERR(handle);
1073                         mlog_errno(status);
1074                         goto bail_unlock;
1075                 }
1076                 status = vfs_dq_transfer(inode, attr) ? -EDQUOT : 0;
1077                 if (status < 0)
1078                         goto bail_commit;
1079         } else {
1080                 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1081                 if (IS_ERR(handle)) {
1082                         status = PTR_ERR(handle);
1083                         mlog_errno(status);
1084                         goto bail_unlock;
1085                 }
1086         }
1087
1088         /*
1089          * This will intentionally not wind up calling vmtruncate(),
1090          * since all the work for a size change has been done above.
1091          * Otherwise, we could get into problems with truncate as
1092          * ip_alloc_sem is used there to protect against i_size
1093          * changes.
1094          */
1095         status = inode_setattr(inode, attr);
1096         if (status < 0) {
1097                 mlog_errno(status);
1098                 goto bail_commit;
1099         }
1100
1101         status = ocfs2_mark_inode_dirty(handle, inode, bh);
1102         if (status < 0)
1103                 mlog_errno(status);
1104
1105 bail_commit:
1106         ocfs2_commit_trans(osb, handle);
1107 bail_unlock:
1108         ocfs2_inode_unlock(inode, 1);
1109 bail_unlock_rw:
1110         if (size_change)
1111                 ocfs2_rw_unlock(inode, 1);
1112 bail:
1113         brelse(bh);
1114
1115         /* Release quota pointers in case we acquired them */
1116         for (qtype = 0; qtype < MAXQUOTAS; qtype++) {
1117                 dqput(transfer_to[qtype]);
1118                 dqput(transfer_from[qtype]);
1119         }
1120
1121         if (!status && attr->ia_valid & ATTR_MODE) {
1122                 status = ocfs2_acl_chmod(inode);
1123                 if (status < 0)
1124                         mlog_errno(status);
1125         }
1126
1127         mlog_exit(status);
1128         return status;
1129 }
1130
1131 int ocfs2_getattr(struct vfsmount *mnt,
1132                   struct dentry *dentry,
1133                   struct kstat *stat)
1134 {
1135         struct inode *inode = dentry->d_inode;
1136         struct super_block *sb = dentry->d_inode->i_sb;
1137         struct ocfs2_super *osb = sb->s_fs_info;
1138         int err;
1139
1140         mlog_entry_void();
1141
1142         err = ocfs2_inode_revalidate(dentry);
1143         if (err) {
1144                 if (err != -ENOENT)
1145                         mlog_errno(err);
1146                 goto bail;
1147         }
1148
1149         generic_fillattr(inode, stat);
1150
1151         /* We set the blksize from the cluster size for performance */
1152         stat->blksize = osb->s_clustersize;
1153
1154 bail:
1155         mlog_exit(err);
1156
1157         return err;
1158 }
1159
1160 int ocfs2_permission(struct inode *inode, int mask)
1161 {
1162         int ret;
1163
1164         mlog_entry_void();
1165
1166         ret = ocfs2_inode_lock(inode, NULL, 0);
1167         if (ret) {
1168                 if (ret != -ENOENT)
1169                         mlog_errno(ret);
1170                 goto out;
1171         }
1172
1173         ret = generic_permission(inode, mask, ocfs2_check_acl);
1174
1175         ocfs2_inode_unlock(inode, 0);
1176 out:
1177         mlog_exit(ret);
1178         return ret;
1179 }
1180
1181 static int __ocfs2_write_remove_suid(struct inode *inode,
1182                                      struct buffer_head *bh)
1183 {
1184         int ret;
1185         handle_t *handle;
1186         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1187         struct ocfs2_dinode *di;
1188
1189         mlog_entry("(Inode %llu, mode 0%o)\n",
1190                    (unsigned long long)OCFS2_I(inode)->ip_blkno, inode->i_mode);
1191
1192         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1193         if (IS_ERR(handle)) {
1194                 ret = PTR_ERR(handle);
1195                 mlog_errno(ret);
1196                 goto out;
1197         }
1198
1199         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), bh,
1200                                       OCFS2_JOURNAL_ACCESS_WRITE);
1201         if (ret < 0) {
1202                 mlog_errno(ret);
1203                 goto out_trans;
1204         }
1205
1206         inode->i_mode &= ~S_ISUID;
1207         if ((inode->i_mode & S_ISGID) && (inode->i_mode & S_IXGRP))
1208                 inode->i_mode &= ~S_ISGID;
1209
1210         di = (struct ocfs2_dinode *) bh->b_data;
1211         di->i_mode = cpu_to_le16(inode->i_mode);
1212
1213         ret = ocfs2_journal_dirty(handle, bh);
1214         if (ret < 0)
1215                 mlog_errno(ret);
1216
1217 out_trans:
1218         ocfs2_commit_trans(osb, handle);
1219 out:
1220         mlog_exit(ret);
1221         return ret;
1222 }
1223
1224 /*
1225  * Will look for holes and unwritten extents in the range starting at
1226  * pos for count bytes (inclusive).
1227  */
1228 static int ocfs2_check_range_for_holes(struct inode *inode, loff_t pos,
1229                                        size_t count)
1230 {
1231         int ret = 0;
1232         unsigned int extent_flags;
1233         u32 cpos, clusters, extent_len, phys_cpos;
1234         struct super_block *sb = inode->i_sb;
1235
1236         cpos = pos >> OCFS2_SB(sb)->s_clustersize_bits;
1237         clusters = ocfs2_clusters_for_bytes(sb, pos + count) - cpos;
1238
1239         while (clusters) {
1240                 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos, &extent_len,
1241                                          &extent_flags);
1242                 if (ret < 0) {
1243                         mlog_errno(ret);
1244                         goto out;
1245                 }
1246
1247                 if (phys_cpos == 0 || (extent_flags & OCFS2_EXT_UNWRITTEN)) {
1248                         ret = 1;
1249                         break;
1250                 }
1251
1252                 if (extent_len > clusters)
1253                         extent_len = clusters;
1254
1255                 clusters -= extent_len;
1256                 cpos += extent_len;
1257         }
1258 out:
1259         return ret;
1260 }
1261
1262 static int ocfs2_write_remove_suid(struct inode *inode)
1263 {
1264         int ret;
1265         struct buffer_head *bh = NULL;
1266
1267         ret = ocfs2_read_inode_block(inode, &bh);
1268         if (ret < 0) {
1269                 mlog_errno(ret);
1270                 goto out;
1271         }
1272
1273         ret =  __ocfs2_write_remove_suid(inode, bh);
1274 out:
1275         brelse(bh);
1276         return ret;
1277 }
1278
1279 /*
1280  * Allocate enough extents to cover the region starting at byte offset
1281  * start for len bytes. Existing extents are skipped, any extents
1282  * added are marked as "unwritten".
1283  */
1284 static int ocfs2_allocate_unwritten_extents(struct inode *inode,
1285                                             u64 start, u64 len)
1286 {
1287         int ret;
1288         u32 cpos, phys_cpos, clusters, alloc_size;
1289         u64 end = start + len;
1290         struct buffer_head *di_bh = NULL;
1291
1292         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1293                 ret = ocfs2_read_inode_block(inode, &di_bh);
1294                 if (ret) {
1295                         mlog_errno(ret);
1296                         goto out;
1297                 }
1298
1299                 /*
1300                  * Nothing to do if the requested reservation range
1301                  * fits within the inode.
1302                  */
1303                 if (ocfs2_size_fits_inline_data(di_bh, end))
1304                         goto out;
1305
1306                 ret = ocfs2_convert_inline_data_to_extents(inode, di_bh);
1307                 if (ret) {
1308                         mlog_errno(ret);
1309                         goto out;
1310                 }
1311         }
1312
1313         /*
1314          * We consider both start and len to be inclusive.
1315          */
1316         cpos = start >> OCFS2_SB(inode->i_sb)->s_clustersize_bits;
1317         clusters = ocfs2_clusters_for_bytes(inode->i_sb, start + len);
1318         clusters -= cpos;
1319
1320         while (clusters) {
1321                 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos,
1322                                          &alloc_size, NULL);
1323                 if (ret) {
1324                         mlog_errno(ret);
1325                         goto out;
1326                 }
1327
1328                 /*
1329                  * Hole or existing extent len can be arbitrary, so
1330                  * cap it to our own allocation request.
1331                  */
1332                 if (alloc_size > clusters)
1333                         alloc_size = clusters;
1334
1335                 if (phys_cpos) {
1336                         /*
1337                          * We already have an allocation at this
1338                          * region so we can safely skip it.
1339                          */
1340                         goto next;
1341                 }
1342
1343                 ret = __ocfs2_extend_allocation(inode, cpos, alloc_size, 1);
1344                 if (ret) {
1345                         if (ret != -ENOSPC)
1346                                 mlog_errno(ret);
1347                         goto out;
1348                 }
1349
1350 next:
1351                 cpos += alloc_size;
1352                 clusters -= alloc_size;
1353         }
1354
1355         ret = 0;
1356 out:
1357
1358         brelse(di_bh);
1359         return ret;
1360 }
1361
1362 /*
1363  * Truncate a byte range, avoiding pages within partial clusters. This
1364  * preserves those pages for the zeroing code to write to.
1365  */
1366 static void ocfs2_truncate_cluster_pages(struct inode *inode, u64 byte_start,
1367                                          u64 byte_len)
1368 {
1369         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1370         loff_t start, end;
1371         struct address_space *mapping = inode->i_mapping;
1372
1373         start = (loff_t)ocfs2_align_bytes_to_clusters(inode->i_sb, byte_start);
1374         end = byte_start + byte_len;
1375         end = end & ~(osb->s_clustersize - 1);
1376
1377         if (start < end) {
1378                 unmap_mapping_range(mapping, start, end - start, 0);
1379                 truncate_inode_pages_range(mapping, start, end - 1);
1380         }
1381 }
1382
1383 static int ocfs2_zero_partial_clusters(struct inode *inode,
1384                                        u64 start, u64 len)
1385 {
1386         int ret = 0;
1387         u64 tmpend, end = start + len;
1388         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1389         unsigned int csize = osb->s_clustersize;
1390         handle_t *handle;
1391
1392         /*
1393          * The "start" and "end" values are NOT necessarily part of
1394          * the range whose allocation is being deleted. Rather, this
1395          * is what the user passed in with the request. We must zero
1396          * partial clusters here. There's no need to worry about
1397          * physical allocation - the zeroing code knows to skip holes.
1398          */
1399         mlog(0, "byte start: %llu, end: %llu\n",
1400              (unsigned long long)start, (unsigned long long)end);
1401
1402         /*
1403          * If both edges are on a cluster boundary then there's no
1404          * zeroing required as the region is part of the allocation to
1405          * be truncated.
1406          */
1407         if ((start & (csize - 1)) == 0 && (end & (csize - 1)) == 0)
1408                 goto out;
1409
1410         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1411         if (IS_ERR(handle)) {
1412                 ret = PTR_ERR(handle);
1413                 mlog_errno(ret);
1414                 goto out;
1415         }
1416
1417         /*
1418          * We want to get the byte offset of the end of the 1st cluster.
1419          */
1420         tmpend = (u64)osb->s_clustersize + (start & ~(osb->s_clustersize - 1));
1421         if (tmpend > end)
1422                 tmpend = end;
1423
1424         mlog(0, "1st range: start: %llu, tmpend: %llu\n",
1425              (unsigned long long)start, (unsigned long long)tmpend);
1426
1427         ret = ocfs2_zero_range_for_truncate(inode, handle, start, tmpend);
1428         if (ret)
1429                 mlog_errno(ret);
1430
1431         if (tmpend < end) {
1432                 /*
1433                  * This may make start and end equal, but the zeroing
1434                  * code will skip any work in that case so there's no
1435                  * need to catch it up here.
1436                  */
1437                 start = end & ~(osb->s_clustersize - 1);
1438
1439                 mlog(0, "2nd range: start: %llu, end: %llu\n",
1440                      (unsigned long long)start, (unsigned long long)end);
1441
1442                 ret = ocfs2_zero_range_for_truncate(inode, handle, start, end);
1443                 if (ret)
1444                         mlog_errno(ret);
1445         }
1446
1447         ocfs2_commit_trans(osb, handle);
1448 out:
1449         return ret;
1450 }
1451
1452 static int ocfs2_remove_inode_range(struct inode *inode,
1453                                     struct buffer_head *di_bh, u64 byte_start,
1454                                     u64 byte_len)
1455 {
1456         int ret = 0;
1457         u32 trunc_start, trunc_len, cpos, phys_cpos, alloc_size;
1458         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1459         struct ocfs2_cached_dealloc_ctxt dealloc;
1460         struct address_space *mapping = inode->i_mapping;
1461         struct ocfs2_extent_tree et;
1462
1463         ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), di_bh);
1464         ocfs2_init_dealloc_ctxt(&dealloc);
1465
1466         if (byte_len == 0)
1467                 return 0;
1468
1469         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1470                 ret = ocfs2_truncate_inline(inode, di_bh, byte_start,
1471                                             byte_start + byte_len, 0);
1472                 if (ret) {
1473                         mlog_errno(ret);
1474                         goto out;
1475                 }
1476                 /*
1477                  * There's no need to get fancy with the page cache
1478                  * truncate of an inline-data inode. We're talking
1479                  * about less than a page here, which will be cached
1480                  * in the dinode buffer anyway.
1481                  */
1482                 unmap_mapping_range(mapping, 0, 0, 0);
1483                 truncate_inode_pages(mapping, 0);
1484                 goto out;
1485         }
1486
1487         trunc_start = ocfs2_clusters_for_bytes(osb->sb, byte_start);
1488         trunc_len = (byte_start + byte_len) >> osb->s_clustersize_bits;
1489         if (trunc_len >= trunc_start)
1490                 trunc_len -= trunc_start;
1491         else
1492                 trunc_len = 0;
1493
1494         mlog(0, "Inode: %llu, start: %llu, len: %llu, cstart: %u, clen: %u\n",
1495              (unsigned long long)OCFS2_I(inode)->ip_blkno,
1496              (unsigned long long)byte_start,
1497              (unsigned long long)byte_len, trunc_start, trunc_len);
1498
1499         ret = ocfs2_zero_partial_clusters(inode, byte_start, byte_len);
1500         if (ret) {
1501                 mlog_errno(ret);
1502                 goto out;
1503         }
1504
1505         cpos = trunc_start;
1506         while (trunc_len) {
1507                 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos,
1508                                          &alloc_size, NULL);
1509                 if (ret) {
1510                         mlog_errno(ret);
1511                         goto out;
1512                 }
1513
1514                 if (alloc_size > trunc_len)
1515                         alloc_size = trunc_len;
1516
1517                 /* Only do work for non-holes */
1518                 if (phys_cpos != 0) {
1519                         ret = ocfs2_remove_btree_range(inode, &et, cpos,
1520                                                        phys_cpos, alloc_size,
1521                                                        &dealloc);
1522                         if (ret) {
1523                                 mlog_errno(ret);
1524                                 goto out;
1525                         }
1526                 }
1527
1528                 cpos += alloc_size;
1529                 trunc_len -= alloc_size;
1530         }
1531
1532         ocfs2_truncate_cluster_pages(inode, byte_start, byte_len);
1533
1534 out:
1535         ocfs2_schedule_truncate_log_flush(osb, 1);
1536         ocfs2_run_deallocs(osb, &dealloc);
1537
1538         return ret;
1539 }
1540
1541 /*
1542  * Parts of this function taken from xfs_change_file_space()
1543  */
1544 static int __ocfs2_change_file_space(struct file *file, struct inode *inode,
1545                                      loff_t f_pos, unsigned int cmd,
1546                                      struct ocfs2_space_resv *sr,
1547                                      int change_size)
1548 {
1549         int ret;
1550         s64 llen;
1551         loff_t size;
1552         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1553         struct buffer_head *di_bh = NULL;
1554         handle_t *handle;
1555         unsigned long long max_off = inode->i_sb->s_maxbytes;
1556
1557         if (ocfs2_is_hard_readonly(osb) || ocfs2_is_soft_readonly(osb))
1558                 return -EROFS;
1559
1560         mutex_lock(&inode->i_mutex);
1561
1562         /*
1563          * This prevents concurrent writes on other nodes
1564          */
1565         ret = ocfs2_rw_lock(inode, 1);
1566         if (ret) {
1567                 mlog_errno(ret);
1568                 goto out;
1569         }
1570
1571         ret = ocfs2_inode_lock(inode, &di_bh, 1);
1572         if (ret) {
1573                 mlog_errno(ret);
1574                 goto out_rw_unlock;
1575         }
1576
1577         if (inode->i_flags & (S_IMMUTABLE|S_APPEND)) {
1578                 ret = -EPERM;
1579                 goto out_inode_unlock;
1580         }
1581
1582         switch (sr->l_whence) {
1583         case 0: /*SEEK_SET*/
1584                 break;
1585         case 1: /*SEEK_CUR*/
1586                 sr->l_start += f_pos;
1587                 break;
1588         case 2: /*SEEK_END*/
1589                 sr->l_start += i_size_read(inode);
1590                 break;
1591         default:
1592                 ret = -EINVAL;
1593                 goto out_inode_unlock;
1594         }
1595         sr->l_whence = 0;
1596
1597         llen = sr->l_len > 0 ? sr->l_len - 1 : sr->l_len;
1598
1599         if (sr->l_start < 0
1600             || sr->l_start > max_off
1601             || (sr->l_start + llen) < 0
1602             || (sr->l_start + llen) > max_off) {
1603                 ret = -EINVAL;
1604                 goto out_inode_unlock;
1605         }
1606         size = sr->l_start + sr->l_len;
1607
1608         if (cmd == OCFS2_IOC_RESVSP || cmd == OCFS2_IOC_RESVSP64) {
1609                 if (sr->l_len <= 0) {
1610                         ret = -EINVAL;
1611                         goto out_inode_unlock;
1612                 }
1613         }
1614
1615         if (file && should_remove_suid(file->f_path.dentry)) {
1616                 ret = __ocfs2_write_remove_suid(inode, di_bh);
1617                 if (ret) {
1618                         mlog_errno(ret);
1619                         goto out_inode_unlock;
1620                 }
1621         }
1622
1623         down_write(&OCFS2_I(inode)->ip_alloc_sem);
1624         switch (cmd) {
1625         case OCFS2_IOC_RESVSP:
1626         case OCFS2_IOC_RESVSP64:
1627                 /*
1628                  * This takes unsigned offsets, but the signed ones we
1629                  * pass have been checked against overflow above.
1630                  */
1631                 ret = ocfs2_allocate_unwritten_extents(inode, sr->l_start,
1632                                                        sr->l_len);
1633                 break;
1634         case OCFS2_IOC_UNRESVSP:
1635         case OCFS2_IOC_UNRESVSP64:
1636                 ret = ocfs2_remove_inode_range(inode, di_bh, sr->l_start,
1637                                                sr->l_len);
1638                 break;
1639         default:
1640                 ret = -EINVAL;
1641         }
1642         up_write(&OCFS2_I(inode)->ip_alloc_sem);
1643         if (ret) {
1644                 mlog_errno(ret);
1645                 goto out_inode_unlock;
1646         }
1647
1648         /*
1649          * We update c/mtime for these changes
1650          */
1651         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1652         if (IS_ERR(handle)) {
1653                 ret = PTR_ERR(handle);
1654                 mlog_errno(ret);
1655                 goto out_inode_unlock;
1656         }
1657
1658         if (change_size && i_size_read(inode) < size)
1659                 i_size_write(inode, size);
1660
1661         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
1662         ret = ocfs2_mark_inode_dirty(handle, inode, di_bh);
1663         if (ret < 0)
1664                 mlog_errno(ret);
1665
1666         ocfs2_commit_trans(osb, handle);
1667
1668 out_inode_unlock:
1669         brelse(di_bh);
1670         ocfs2_inode_unlock(inode, 1);
1671 out_rw_unlock:
1672         ocfs2_rw_unlock(inode, 1);
1673
1674 out:
1675         mutex_unlock(&inode->i_mutex);
1676         return ret;
1677 }
1678
1679 int ocfs2_change_file_space(struct file *file, unsigned int cmd,
1680                             struct ocfs2_space_resv *sr)
1681 {
1682         struct inode *inode = file->f_path.dentry->d_inode;
1683         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1684
1685         if ((cmd == OCFS2_IOC_RESVSP || cmd == OCFS2_IOC_RESVSP64) &&
1686             !ocfs2_writes_unwritten_extents(osb))
1687                 return -ENOTTY;
1688         else if ((cmd == OCFS2_IOC_UNRESVSP || cmd == OCFS2_IOC_UNRESVSP64) &&
1689                  !ocfs2_sparse_alloc(osb))
1690                 return -ENOTTY;
1691
1692         if (!S_ISREG(inode->i_mode))
1693                 return -EINVAL;
1694
1695         if (!(file->f_mode & FMODE_WRITE))
1696                 return -EBADF;
1697
1698         return __ocfs2_change_file_space(file, inode, file->f_pos, cmd, sr, 0);
1699 }
1700
1701 static long ocfs2_fallocate(struct inode *inode, int mode, loff_t offset,
1702                             loff_t len)
1703 {
1704         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1705         struct ocfs2_space_resv sr;
1706         int change_size = 1;
1707
1708         if (!ocfs2_writes_unwritten_extents(osb))
1709                 return -EOPNOTSUPP;
1710
1711         if (S_ISDIR(inode->i_mode))
1712                 return -ENODEV;
1713
1714         if (mode & FALLOC_FL_KEEP_SIZE)
1715                 change_size = 0;
1716
1717         sr.l_whence = 0;
1718         sr.l_start = (s64)offset;
1719         sr.l_len = (s64)len;
1720
1721         return __ocfs2_change_file_space(NULL, inode, offset,
1722                                          OCFS2_IOC_RESVSP64, &sr, change_size);
1723 }
1724
1725 int ocfs2_check_range_for_refcount(struct inode *inode, loff_t pos,
1726                                    size_t count)
1727 {
1728         int ret = 0;
1729         unsigned int extent_flags;
1730         u32 cpos, clusters, extent_len, phys_cpos;
1731         struct super_block *sb = inode->i_sb;
1732
1733         if (!ocfs2_refcount_tree(OCFS2_SB(inode->i_sb)) ||
1734             !(OCFS2_I(inode)->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL) ||
1735             OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
1736                 return 0;
1737
1738         cpos = pos >> OCFS2_SB(sb)->s_clustersize_bits;
1739         clusters = ocfs2_clusters_for_bytes(sb, pos + count) - cpos;
1740
1741         while (clusters) {
1742                 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos, &extent_len,
1743                                          &extent_flags);
1744                 if (ret < 0) {
1745                         mlog_errno(ret);
1746                         goto out;
1747                 }
1748
1749                 if (phys_cpos && (extent_flags & OCFS2_EXT_REFCOUNTED)) {
1750                         ret = 1;
1751                         break;
1752                 }
1753
1754                 if (extent_len > clusters)
1755                         extent_len = clusters;
1756
1757                 clusters -= extent_len;
1758                 cpos += extent_len;
1759         }
1760 out:
1761         return ret;
1762 }
1763
1764 static int ocfs2_prepare_inode_for_refcount(struct inode *inode,
1765                                             loff_t pos, size_t count,
1766                                             int *meta_level)
1767 {
1768         int ret;
1769         struct buffer_head *di_bh = NULL;
1770         u32 cpos = pos >> OCFS2_SB(inode->i_sb)->s_clustersize_bits;
1771         u32 clusters =
1772                 ocfs2_clusters_for_bytes(inode->i_sb, pos + count) - cpos;
1773
1774         ret = ocfs2_inode_lock(inode, &di_bh, 1);
1775         if (ret) {
1776                 mlog_errno(ret);
1777                 goto out;
1778         }
1779
1780         *meta_level = 1;
1781
1782         ret = ocfs2_refcount_cow(inode, di_bh, cpos, clusters, UINT_MAX);
1783         if (ret)
1784                 mlog_errno(ret);
1785 out:
1786         brelse(di_bh);
1787         return ret;
1788 }
1789
1790 static int ocfs2_prepare_inode_for_write(struct dentry *dentry,
1791                                          loff_t *ppos,
1792                                          size_t count,
1793                                          int appending,
1794                                          int *direct_io,
1795                                          int *has_refcount)
1796 {
1797         int ret = 0, meta_level = 0;
1798         struct inode *inode = dentry->d_inode;
1799         loff_t saved_pos, end;
1800
1801         /*
1802          * We start with a read level meta lock and only jump to an ex
1803          * if we need to make modifications here.
1804          */
1805         for(;;) {
1806                 ret = ocfs2_inode_lock(inode, NULL, meta_level);
1807                 if (ret < 0) {
1808                         meta_level = -1;
1809                         mlog_errno(ret);
1810                         goto out;
1811                 }
1812
1813                 /* Clear suid / sgid if necessary. We do this here
1814                  * instead of later in the write path because
1815                  * remove_suid() calls ->setattr without any hint that
1816                  * we may have already done our cluster locking. Since
1817                  * ocfs2_setattr() *must* take cluster locks to
1818                  * proceeed, this will lead us to recursively lock the
1819                  * inode. There's also the dinode i_size state which
1820                  * can be lost via setattr during extending writes (we
1821                  * set inode->i_size at the end of a write. */
1822                 if (should_remove_suid(dentry)) {
1823                         if (meta_level == 0) {
1824                                 ocfs2_inode_unlock(inode, meta_level);
1825                                 meta_level = 1;
1826                                 continue;
1827                         }
1828
1829                         ret = ocfs2_write_remove_suid(inode);
1830                         if (ret < 0) {
1831                                 mlog_errno(ret);
1832                                 goto out_unlock;
1833                         }
1834                 }
1835
1836                 /* work on a copy of ppos until we're sure that we won't have
1837                  * to recalculate it due to relocking. */
1838                 if (appending) {
1839                         saved_pos = i_size_read(inode);
1840                         mlog(0, "O_APPEND: inode->i_size=%llu\n", saved_pos);
1841                 } else {
1842                         saved_pos = *ppos;
1843                 }
1844
1845                 end = saved_pos + count;
1846
1847                 ret = ocfs2_check_range_for_refcount(inode, saved_pos, count);
1848                 if (ret == 1) {
1849                         ocfs2_inode_unlock(inode, meta_level);
1850                         meta_level = -1;
1851
1852                         ret = ocfs2_prepare_inode_for_refcount(inode,
1853                                                                saved_pos,
1854                                                                count,
1855                                                                &meta_level);
1856                         if (has_refcount)
1857                                 *has_refcount = 1;
1858                 }
1859
1860                 if (ret < 0) {
1861                         mlog_errno(ret);
1862                         goto out_unlock;
1863                 }
1864
1865                 /*
1866                  * Skip the O_DIRECT checks if we don't need
1867                  * them.
1868                  */
1869                 if (!direct_io || !(*direct_io))
1870                         break;
1871
1872                 /*
1873                  * There's no sane way to do direct writes to an inode
1874                  * with inline data.
1875                  */
1876                 if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1877                         *direct_io = 0;
1878                         break;
1879                 }
1880
1881                 if (has_refcount && *has_refcount == 1) {
1882                         *direct_io = 0;
1883                         break;
1884                 }
1885                 /*
1886                  * Allowing concurrent direct writes means
1887                  * i_size changes wouldn't be synchronized, so
1888                  * one node could wind up truncating another
1889                  * nodes writes.
1890                  */
1891                 if (end > i_size_read(inode)) {
1892                         *direct_io = 0;
1893                         break;
1894                 }
1895
1896                 /*
1897                  * We don't fill holes during direct io, so
1898                  * check for them here. If any are found, the
1899                  * caller will have to retake some cluster
1900                  * locks and initiate the io as buffered.
1901                  */
1902                 ret = ocfs2_check_range_for_holes(inode, saved_pos, count);
1903                 if (ret == 1) {
1904                         *direct_io = 0;
1905                         ret = 0;
1906                 } else if (ret < 0)
1907                         mlog_errno(ret);
1908                 break;
1909         }
1910
1911         if (appending)
1912                 *ppos = saved_pos;
1913
1914 out_unlock:
1915         if (meta_level >= 0)
1916                 ocfs2_inode_unlock(inode, meta_level);
1917
1918 out:
1919         return ret;
1920 }
1921
1922 static ssize_t ocfs2_file_aio_write(struct kiocb *iocb,
1923                                     const struct iovec *iov,
1924                                     unsigned long nr_segs,
1925                                     loff_t pos)
1926 {
1927         int ret, direct_io, appending, rw_level, have_alloc_sem  = 0;
1928         int can_do_direct, has_refcount = 0;
1929         ssize_t written = 0;
1930         size_t ocount;          /* original count */
1931         size_t count;           /* after file limit checks */
1932         loff_t old_size, *ppos = &iocb->ki_pos;
1933         u32 old_clusters;
1934         struct file *file = iocb->ki_filp;
1935         struct inode *inode = file->f_path.dentry->d_inode;
1936         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1937
1938         mlog_entry("(0x%p, %u, '%.*s')\n", file,
1939                    (unsigned int)nr_segs,
1940                    file->f_path.dentry->d_name.len,
1941                    file->f_path.dentry->d_name.name);
1942
1943         if (iocb->ki_left == 0)
1944                 return 0;
1945
1946         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
1947
1948         appending = file->f_flags & O_APPEND ? 1 : 0;
1949         direct_io = file->f_flags & O_DIRECT ? 1 : 0;
1950
1951         mutex_lock(&inode->i_mutex);
1952
1953 relock:
1954         /* to match setattr's i_mutex -> i_alloc_sem -> rw_lock ordering */
1955         if (direct_io) {
1956                 down_read(&inode->i_alloc_sem);
1957                 have_alloc_sem = 1;
1958         }
1959
1960         /* concurrent O_DIRECT writes are allowed */
1961         rw_level = !direct_io;
1962         ret = ocfs2_rw_lock(inode, rw_level);
1963         if (ret < 0) {
1964                 mlog_errno(ret);
1965                 goto out_sems;
1966         }
1967
1968         can_do_direct = direct_io;
1969         ret = ocfs2_prepare_inode_for_write(file->f_path.dentry, ppos,
1970                                             iocb->ki_left, appending,
1971                                             &can_do_direct, &has_refcount);
1972         if (ret < 0) {
1973                 mlog_errno(ret);
1974                 goto out;
1975         }
1976
1977         /*
1978          * We can't complete the direct I/O as requested, fall back to
1979          * buffered I/O.
1980          */
1981         if (direct_io && !can_do_direct) {
1982                 ocfs2_rw_unlock(inode, rw_level);
1983                 up_read(&inode->i_alloc_sem);
1984
1985                 have_alloc_sem = 0;
1986                 rw_level = -1;
1987
1988                 direct_io = 0;
1989                 goto relock;
1990         }
1991
1992         /*
1993          * To later detect whether a journal commit for sync writes is
1994          * necessary, we sample i_size, and cluster count here.
1995          */
1996         old_size = i_size_read(inode);
1997         old_clusters = OCFS2_I(inode)->ip_clusters;
1998
1999         /* communicate with ocfs2_dio_end_io */
2000         ocfs2_iocb_set_rw_locked(iocb, rw_level);
2001
2002         if (direct_io) {
2003                 ret = generic_segment_checks(iov, &nr_segs, &ocount,
2004                                              VERIFY_READ);
2005                 if (ret)
2006                         goto out_dio;
2007
2008                 count = ocount;
2009                 ret = generic_write_checks(file, ppos, &count,
2010                                            S_ISBLK(inode->i_mode));
2011                 if (ret)
2012                         goto out_dio;
2013
2014                 written = generic_file_direct_write(iocb, iov, &nr_segs, *ppos,
2015                                                     ppos, count, ocount);
2016                 if (written < 0) {
2017                         /*
2018                          * direct write may have instantiated a few
2019                          * blocks outside i_size. Trim these off again.
2020                          * Don't need i_size_read because we hold i_mutex.
2021                          */
2022                         if (*ppos + count > inode->i_size)
2023                                 vmtruncate(inode, inode->i_size);
2024                         ret = written;
2025                         goto out_dio;
2026                 }
2027         } else {
2028                 written = __generic_file_aio_write(iocb, iov, nr_segs, ppos);
2029         }
2030
2031 out_dio:
2032         /* buffered aio wouldn't have proper lock coverage today */
2033         BUG_ON(ret == -EIOCBQUEUED && !(file->f_flags & O_DIRECT));
2034
2035         if ((file->f_flags & O_DSYNC && !direct_io) || IS_SYNC(inode) ||
2036             (file->f_flags & O_DIRECT && has_refcount)) {
2037                 ret = filemap_fdatawrite_range(file->f_mapping, pos,
2038                                                pos + count - 1);
2039                 if (ret < 0)
2040                         written = ret;
2041
2042                 if (!ret && (old_size != i_size_read(inode) ||
2043                     old_clusters != OCFS2_I(inode)->ip_clusters ||
2044                     has_refcount)) {
2045                         ret = jbd2_journal_force_commit(osb->journal->j_journal);
2046                         if (ret < 0)
2047                                 written = ret;
2048                 }
2049
2050                 if (!ret)
2051                         ret = filemap_fdatawait_range(file->f_mapping, pos,
2052                                                       pos + count - 1);
2053         }
2054
2055         /*
2056          * deep in g_f_a_w_n()->ocfs2_direct_IO we pass in a ocfs2_dio_end_io
2057          * function pointer which is called when o_direct io completes so that
2058          * it can unlock our rw lock.  (it's the clustered equivalent of
2059          * i_alloc_sem; protects truncate from racing with pending ios).
2060          * Unfortunately there are error cases which call end_io and others
2061          * that don't.  so we don't have to unlock the rw_lock if either an
2062          * async dio is going to do it in the future or an end_io after an
2063          * error has already done it.
2064          */
2065         if (ret == -EIOCBQUEUED || !ocfs2_iocb_is_rw_locked(iocb)) {
2066                 rw_level = -1;
2067                 have_alloc_sem = 0;
2068         }
2069
2070 out:
2071         if (rw_level != -1)
2072                 ocfs2_rw_unlock(inode, rw_level);
2073
2074 out_sems:
2075         if (have_alloc_sem)
2076                 up_read(&inode->i_alloc_sem);
2077
2078         mutex_unlock(&inode->i_mutex);
2079
2080         if (written)
2081                 ret = written;
2082         mlog_exit(ret);
2083         return ret;
2084 }
2085
2086 static int ocfs2_splice_to_file(struct pipe_inode_info *pipe,
2087                                 struct file *out,
2088                                 struct splice_desc *sd)
2089 {
2090         int ret;
2091
2092         ret = ocfs2_prepare_inode_for_write(out->f_path.dentry, &sd->pos,
2093                                             sd->total_len, 0, NULL, NULL);
2094         if (ret < 0) {
2095                 mlog_errno(ret);
2096                 return ret;
2097         }
2098
2099         return splice_from_pipe_feed(pipe, sd, pipe_to_file);
2100 }
2101
2102 static ssize_t ocfs2_file_splice_write(struct pipe_inode_info *pipe,
2103                                        struct file *out,
2104                                        loff_t *ppos,
2105                                        size_t len,
2106                                        unsigned int flags)
2107 {
2108         int ret;
2109         struct address_space *mapping = out->f_mapping;
2110         struct inode *inode = mapping->host;
2111         struct splice_desc sd = {
2112                 .total_len = len,
2113                 .flags = flags,
2114                 .pos = *ppos,
2115                 .u.file = out,
2116         };
2117
2118         mlog_entry("(0x%p, 0x%p, %u, '%.*s')\n", out, pipe,
2119                    (unsigned int)len,
2120                    out->f_path.dentry->d_name.len,
2121                    out->f_path.dentry->d_name.name);
2122
2123         if (pipe->inode)
2124                 mutex_lock_nested(&pipe->inode->i_mutex, I_MUTEX_PARENT);
2125
2126         splice_from_pipe_begin(&sd);
2127         do {
2128                 ret = splice_from_pipe_next(pipe, &sd);
2129                 if (ret <= 0)
2130                         break;
2131
2132                 mutex_lock_nested(&inode->i_mutex, I_MUTEX_CHILD);
2133                 ret = ocfs2_rw_lock(inode, 1);
2134                 if (ret < 0)
2135                         mlog_errno(ret);
2136                 else {
2137                         ret = ocfs2_splice_to_file(pipe, out, &sd);
2138                         ocfs2_rw_unlock(inode, 1);
2139                 }
2140                 mutex_unlock(&inode->i_mutex);
2141         } while (ret > 0);
2142         splice_from_pipe_end(pipe, &sd);
2143
2144         if (pipe->inode)
2145                 mutex_unlock(&pipe->inode->i_mutex);
2146
2147         if (sd.num_spliced)
2148                 ret = sd.num_spliced;
2149
2150         if (ret > 0) {
2151                 unsigned long nr_pages;
2152                 int err;
2153
2154                 nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
2155
2156                 err = generic_write_sync(out, *ppos, ret);
2157                 if (err)
2158                         ret = err;
2159                 else
2160                         *ppos += ret;
2161
2162                 balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
2163         }
2164
2165         mlog_exit(ret);
2166         return ret;
2167 }
2168
2169 static ssize_t ocfs2_file_splice_read(struct file *in,
2170                                       loff_t *ppos,
2171                                       struct pipe_inode_info *pipe,
2172                                       size_t len,
2173                                       unsigned int flags)
2174 {
2175         int ret = 0, lock_level = 0;
2176         struct inode *inode = in->f_path.dentry->d_inode;
2177
2178         mlog_entry("(0x%p, 0x%p, %u, '%.*s')\n", in, pipe,
2179                    (unsigned int)len,
2180                    in->f_path.dentry->d_name.len,
2181                    in->f_path.dentry->d_name.name);
2182
2183         /*
2184          * See the comment in ocfs2_file_aio_read()
2185          */
2186         ret = ocfs2_inode_lock_atime(inode, in->f_vfsmnt, &lock_level);
2187         if (ret < 0) {
2188                 mlog_errno(ret);
2189                 goto bail;
2190         }
2191         ocfs2_inode_unlock(inode, lock_level);
2192
2193         ret = generic_file_splice_read(in, ppos, pipe, len, flags);
2194
2195 bail:
2196         mlog_exit(ret);
2197         return ret;
2198 }
2199
2200 static ssize_t ocfs2_file_aio_read(struct kiocb *iocb,
2201                                    const struct iovec *iov,
2202                                    unsigned long nr_segs,
2203                                    loff_t pos)
2204 {
2205         int ret = 0, rw_level = -1, have_alloc_sem = 0, lock_level = 0;
2206         struct file *filp = iocb->ki_filp;
2207         struct inode *inode = filp->f_path.dentry->d_inode;
2208
2209         mlog_entry("(0x%p, %u, '%.*s')\n", filp,
2210                    (unsigned int)nr_segs,
2211                    filp->f_path.dentry->d_name.len,
2212                    filp->f_path.dentry->d_name.name);
2213
2214         if (!inode) {
2215                 ret = -EINVAL;
2216                 mlog_errno(ret);
2217                 goto bail;
2218         }
2219
2220         /*
2221          * buffered reads protect themselves in ->readpage().  O_DIRECT reads
2222          * need locks to protect pending reads from racing with truncate.
2223          */
2224         if (filp->f_flags & O_DIRECT) {
2225                 down_read(&inode->i_alloc_sem);
2226                 have_alloc_sem = 1;
2227
2228                 ret = ocfs2_rw_lock(inode, 0);
2229                 if (ret < 0) {
2230                         mlog_errno(ret);
2231                         goto bail;
2232                 }
2233                 rw_level = 0;
2234                 /* communicate with ocfs2_dio_end_io */
2235                 ocfs2_iocb_set_rw_locked(iocb, rw_level);
2236         }
2237
2238         /*
2239          * We're fine letting folks race truncates and extending
2240          * writes with read across the cluster, just like they can
2241          * locally. Hence no rw_lock during read.
2242          *
2243          * Take and drop the meta data lock to update inode fields
2244          * like i_size. This allows the checks down below
2245          * generic_file_aio_read() a chance of actually working.
2246          */
2247         ret = ocfs2_inode_lock_atime(inode, filp->f_vfsmnt, &lock_level);
2248         if (ret < 0) {
2249                 mlog_errno(ret);
2250                 goto bail;
2251         }
2252         ocfs2_inode_unlock(inode, lock_level);
2253
2254         ret = generic_file_aio_read(iocb, iov, nr_segs, iocb->ki_pos);
2255         if (ret == -EINVAL)
2256                 mlog(0, "generic_file_aio_read returned -EINVAL\n");
2257
2258         /* buffered aio wouldn't have proper lock coverage today */
2259         BUG_ON(ret == -EIOCBQUEUED && !(filp->f_flags & O_DIRECT));
2260
2261         /* see ocfs2_file_aio_write */
2262         if (ret == -EIOCBQUEUED || !ocfs2_iocb_is_rw_locked(iocb)) {
2263                 rw_level = -1;
2264                 have_alloc_sem = 0;
2265         }
2266
2267 bail:
2268         if (have_alloc_sem)
2269                 up_read(&inode->i_alloc_sem);
2270         if (rw_level != -1)
2271                 ocfs2_rw_unlock(inode, rw_level);
2272         mlog_exit(ret);
2273
2274         return ret;
2275 }
2276
2277 const struct inode_operations ocfs2_file_iops = {
2278         .setattr        = ocfs2_setattr,
2279         .getattr        = ocfs2_getattr,
2280         .permission     = ocfs2_permission,
2281         .setxattr       = generic_setxattr,
2282         .getxattr       = generic_getxattr,
2283         .listxattr      = ocfs2_listxattr,
2284         .removexattr    = generic_removexattr,
2285         .fallocate      = ocfs2_fallocate,
2286         .fiemap         = ocfs2_fiemap,
2287 };
2288
2289 const struct inode_operations ocfs2_special_file_iops = {
2290         .setattr        = ocfs2_setattr,
2291         .getattr        = ocfs2_getattr,
2292         .permission     = ocfs2_permission,
2293 };
2294
2295 /*
2296  * Other than ->lock, keep ocfs2_fops and ocfs2_dops in sync with
2297  * ocfs2_fops_no_plocks and ocfs2_dops_no_plocks!
2298  */
2299 const struct file_operations ocfs2_fops = {
2300         .llseek         = generic_file_llseek,
2301         .read           = do_sync_read,
2302         .write          = do_sync_write,
2303         .mmap           = ocfs2_mmap,
2304         .fsync          = ocfs2_sync_file,
2305         .release        = ocfs2_file_release,
2306         .open           = ocfs2_file_open,
2307         .aio_read       = ocfs2_file_aio_read,
2308         .aio_write      = ocfs2_file_aio_write,
2309         .unlocked_ioctl = ocfs2_ioctl,
2310 #ifdef CONFIG_COMPAT
2311         .compat_ioctl   = ocfs2_compat_ioctl,
2312 #endif
2313         .lock           = ocfs2_lock,
2314         .flock          = ocfs2_flock,
2315         .splice_read    = ocfs2_file_splice_read,
2316         .splice_write   = ocfs2_file_splice_write,
2317 };
2318
2319 const struct file_operations ocfs2_dops = {
2320         .llseek         = generic_file_llseek,
2321         .read           = generic_read_dir,
2322         .readdir        = ocfs2_readdir,
2323         .fsync          = ocfs2_sync_file,
2324         .release        = ocfs2_dir_release,
2325         .open           = ocfs2_dir_open,
2326         .unlocked_ioctl = ocfs2_ioctl,
2327 #ifdef CONFIG_COMPAT
2328         .compat_ioctl   = ocfs2_compat_ioctl,
2329 #endif
2330         .lock           = ocfs2_lock,
2331         .flock          = ocfs2_flock,
2332 };
2333
2334 /*
2335  * POSIX-lockless variants of our file_operations.
2336  *
2337  * These will be used if the underlying cluster stack does not support
2338  * posix file locking, if the user passes the "localflocks" mount
2339  * option, or if we have a local-only fs.
2340  *
2341  * ocfs2_flock is in here because all stacks handle UNIX file locks,
2342  * so we still want it in the case of no stack support for
2343  * plocks. Internally, it will do the right thing when asked to ignore
2344  * the cluster.
2345  */
2346 const struct file_operations ocfs2_fops_no_plocks = {
2347         .llseek         = generic_file_llseek,
2348         .read           = do_sync_read,
2349         .write          = do_sync_write,
2350         .mmap           = ocfs2_mmap,
2351         .fsync          = ocfs2_sync_file,
2352         .release        = ocfs2_file_release,
2353         .open           = ocfs2_file_open,
2354         .aio_read       = ocfs2_file_aio_read,
2355         .aio_write      = ocfs2_file_aio_write,
2356         .unlocked_ioctl = ocfs2_ioctl,
2357 #ifdef CONFIG_COMPAT
2358         .compat_ioctl   = ocfs2_compat_ioctl,
2359 #endif
2360         .flock          = ocfs2_flock,
2361         .splice_read    = ocfs2_file_splice_read,
2362         .splice_write   = ocfs2_file_splice_write,
2363 };
2364
2365 const struct file_operations ocfs2_dops_no_plocks = {
2366         .llseek         = generic_file_llseek,
2367         .read           = generic_read_dir,
2368         .readdir        = ocfs2_readdir,
2369         .fsync          = ocfs2_sync_file,
2370         .release        = ocfs2_dir_release,
2371         .open           = ocfs2_dir_open,
2372         .unlocked_ioctl = ocfs2_ioctl,
2373 #ifdef CONFIG_COMPAT
2374         .compat_ioctl   = ocfs2_compat_ioctl,
2375 #endif
2376         .flock          = ocfs2_flock,
2377 };