fs/ocfs2/file.c

   1 /* -*- mode: c; c-basic-offset: 8; -*-
   2  * vim: noexpandtab sw=8 ts=8 sts=0:
   3  *
   4  * file.c
   5  *
   6  * File open, close, extend, truncate
   7  *
   8  * Copyright (C) 2002, 2004 Oracle.  All rights reserved.
   9  *
  10  * This program is free software; you can redistribute it and/or
  11  * modify it under the terms of the GNU General Public
  12  * License as published by the Free Software Foundation; either
  13  * version 2 of the License, or (at your option) any later version.
  14  *
  15  * This program is distributed in the hope that it will be useful,
  16  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  17  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  18  * General Public License for more details.
  19  *
  20  * You should have received a copy of the GNU General Public
  21  * License along with this program; if not, write to the
  22  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  23  * Boston, MA 021110-1307, USA.
  24  */
  25
  26 #include <linux/capability.h>
  27 #include <linux/fs.h>
  28 #include <linux/types.h>
  29 #include <linux/slab.h>
  30 #include <linux/highmem.h>
  31 #include <linux/pagemap.h>
  32 #include <linux/uio.h>
  33 #include <linux/sched.h>
  34 #include <linux/splice.h>
  35 #include <linux/mount.h>
  36 #include <linux/writeback.h>
  37 #include <linux/falloc.h>
  38 #include <linux/quotaops.h>
  39
  40 #define MLOG_MASK_PREFIX ML_INODE
  41 #include <cluster/masklog.h>
  42
  43 #include "ocfs2.h"
  44
  45 #include "alloc.h"
  46 #include "aops.h"
  47 #include "dir.h"
  48 #include "dlmglue.h"
  49 #include "extent_map.h"
  50 #include "file.h"
  51 #include "sysfile.h"
  52 #include "inode.h"
  53 #include "ioctl.h"
  54 #include "journal.h"
  55 #include "locks.h"
  56 #include "mmap.h"
  57 #include "suballoc.h"
  58 #include "super.h"
  59 #include "xattr.h"
  60 #include "acl.h"
  61 #include "quota.h"
  62
  63 #include "buffer_head_io.h"
  64
  65 static int ocfs2_sync_inode(struct inode *inode)
  66 {
  67         filemap_fdatawrite(inode->i_mapping);
  68         return sync_mapping_buffers(inode->i_mapping);
  69 }
  70
  71 static int ocfs2_init_file_private(struct inode *inode, struct file *file)
  72 {
  73         struct ocfs2_file_private *fp;
  74
  75         fp = kzalloc(sizeof(struct ocfs2_file_private), GFP_KERNEL);
  76         if (!fp)
  77                 return -ENOMEM;
  78
  79         fp->fp_file = file;
  80         mutex_init(&fp->fp_mutex);
  81         ocfs2_file_lock_res_init(&fp->fp_flock, fp);
  82         file->private_data = fp;
  83
  84         return 0;
  85 }
  86
  87 static void ocfs2_free_file_private(struct inode *inode, struct file *file)
  88 {
  89         struct ocfs2_file_private *fp = file->private_data;
  90         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
  91
  92         if (fp) {
  93                 ocfs2_simple_drop_lockres(osb, &fp->fp_flock);
  94                 ocfs2_lock_res_free(&fp->fp_flock);
  95                 kfree(fp);
  96                 file->private_data = NULL;
  97         }
  98 }
  99
 100 static int ocfs2_file_open(struct inode *inode, struct file *file)
 101 {
 102         int status;
 103         int mode = file->f_flags;
 104         struct ocfs2_inode_info *oi = OCFS2_I(inode);
 105
 106         mlog_entry("(0x%p, 0x%p, '%.*s')\n", inode, file,
 107                    file->f_path.dentry->d_name.len, file->f_path.dentry->d_name.name);
 108
 109         spin_lock(&oi->ip_lock);
 110
 111         /* Check that the inode hasn't been wiped from disk by another
 112          * node. If it hasn't then we're safe as long as we hold the
 113          * spin lock until our increment of open count. */
 114         if (OCFS2_I(inode)->ip_flags & OCFS2_INODE_DELETED) {
 115                 spin_unlock(&oi->ip_lock);
 116
 117                 status = -ENOENT;
 118                 goto leave;
 119         }
 120
 121         if (mode & O_DIRECT)
 122                 oi->ip_flags |= OCFS2_INODE_OPEN_DIRECT;
 123
 124         oi->ip_open_count++;
 125         spin_unlock(&oi->ip_lock);
 126
 127         status = ocfs2_init_file_private(inode, file);
 128         if (status) {
 129                 /*
 130                  * We want to set open count back if we're failing the
 131                  * open.
 132                  */
 133                 spin_lock(&oi->ip_lock);
 134                 oi->ip_open_count--;
 135                 spin_unlock(&oi->ip_lock);
 136         }
 137
 138 leave:
 139         mlog_exit(status);
 140         return status;
 141 }
 142
 143 static int ocfs2_file_release(struct inode *inode, struct file *file)
 144 {
 145         struct ocfs2_inode_info *oi = OCFS2_I(inode);
 146
 147         mlog_entry("(0x%p, 0x%p, '%.*s')\n", inode, file,
 148                        file->f_path.dentry->d_name.len,
 149                        file->f_path.dentry->d_name.name);
 150
 151         spin_lock(&oi->ip_lock);
 152         if (!--oi->ip_open_count)
 153                 oi->ip_flags &= ~OCFS2_INODE_OPEN_DIRECT;
 154         spin_unlock(&oi->ip_lock);
 155
 156         ocfs2_free_file_private(inode, file);
 157
 158         mlog_exit(0);
 159
 160         return 0;
 161 }
 162
 163 static int ocfs2_dir_open(struct inode *inode, struct file *file)
 164 {
 165         return ocfs2_init_file_private(inode, file);
 166 }
 167
 168 static int ocfs2_dir_release(struct inode *inode, struct file *file)
 169 {
 170         ocfs2_free_file_private(inode, file);
 171         return 0;
 172 }
 173
 174 static int ocfs2_sync_file(struct file *file,
 175                            struct dentry *dentry,
 176                            int datasync)
 177 {
 178         int err = 0;
 179         journal_t *journal;
 180         struct inode *inode = dentry->d_inode;
 181         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 182
 183         mlog_entry("(0x%p, 0x%p, %d, '%.*s')\n", file, dentry, datasync,
 184                    dentry->d_name.len, dentry->d_name.name);
 185
 186         err = ocfs2_sync_inode(dentry->d_inode);
 187         if (err)
 188                 goto bail;
 189
 190         journal = osb->journal->j_journal;
 191         err = jbd2_journal_force_commit(journal);
 192
 193 bail:
 194         mlog_exit(err);
 195
 196         return (err < 0) ? -EIO : 0;
 197 }
 198
 199 int ocfs2_should_update_atime(struct inode *inode,
 200                               struct vfsmount *vfsmnt)
 201 {
 202         struct timespec now;
 203         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 204
 205         if (ocfs2_is_hard_readonly(osb) || ocfs2_is_soft_readonly(osb))
 206                 return 0;
 207
 208         if ((inode->i_flags & S_NOATIME) ||
 209             ((inode->i_sb->s_flags & MS_NODIRATIME) && S_ISDIR(inode->i_mode)))
 210                 return 0;
 211
 212         /*
 213          * We can be called with no vfsmnt structure - NFSD will
 214          * sometimes do this.
 215          *
 216          * Note that our action here is different than touch_atime() -
 217          * if we can't tell whether this is a noatime mount, then we
 218          * don't know whether to trust the value of s_atime_quantum.
 219          */
 220         if (vfsmnt == NULL)
 221                 return 0;
 222
 223         if ((vfsmnt->mnt_flags & MNT_NOATIME) ||
 224             ((vfsmnt->mnt_flags & MNT_NODIRATIME) && S_ISDIR(inode->i_mode)))
 225                 return 0;
 226
 227         if (vfsmnt->mnt_flags & MNT_RELATIME) {
 228                 if ((timespec_compare(&inode->i_atime, &inode->i_mtime) <= 0) ||
 229                     (timespec_compare(&inode->i_atime, &inode->i_ctime) <= 0))
 230                         return 1;
 231
 232                 return 0;
 233         }
 234
 235         now = CURRENT_TIME;
 236         if ((now.tv_sec - inode->i_atime.tv_sec <= osb->s_atime_quantum))
 237                 return 0;
 238         else
 239                 return 1;
 240 }
 241
 242 int ocfs2_update_inode_atime(struct inode *inode,
 243                              struct buffer_head *bh)
 244 {
 245         int ret;
 246         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 247         handle_t *handle;
 248         struct ocfs2_dinode *di = (struct ocfs2_dinode *) bh->b_data;
 249
 250         mlog_entry_void();
 251
 252         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
 253         if (IS_ERR(handle)) {
 254                 ret = PTR_ERR(handle);
 255                 mlog_errno(ret);
 256                 goto out;
 257         }
 258
 259         ret = ocfs2_journal_access_di(handle, inode, bh,
 260                                       OCFS2_JOURNAL_ACCESS_WRITE);
 261         if (ret) {
 262                 mlog_errno(ret);
 263                 goto out_commit;
 264         }
 265
 266         /*
 267          * Don't use ocfs2_mark_inode_dirty() here as we don't always
 268          * have i_mutex to guard against concurrent changes to other
 269          * inode fields.
 270          */
 271         inode->i_atime = CURRENT_TIME;
 272         di->i_atime = cpu_to_le64(inode->i_atime.tv_sec);
 273         di->i_atime_nsec = cpu_to_le32(inode->i_atime.tv_nsec);
 274
 275         ret = ocfs2_journal_dirty(handle, bh);
 276         if (ret < 0)
 277                 mlog_errno(ret);
 278
 279 out_commit:
 280         ocfs2_commit_trans(OCFS2_SB(inode->i_sb), handle);
 281 out:
 282         mlog_exit(ret);
 283         return ret;
 284 }
 285
 286 static int ocfs2_set_inode_size(handle_t *handle,
 287                                 struct inode *inode,
 288                                 struct buffer_head *fe_bh,
 289                                 u64 new_i_size)
 290 {
 291         int status;
 292
 293         mlog_entry_void();
 294         i_size_write(inode, new_i_size);
 295         inode->i_blocks = ocfs2_inode_sector_count(inode);
 296         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
 297
 298         status = ocfs2_mark_inode_dirty(handle, inode, fe_bh);
 299         if (status < 0) {
 300                 mlog_errno(status);
 301                 goto bail;
 302         }
 303
 304 bail:
 305         mlog_exit(status);
 306         return status;
 307 }
 308
 309 int ocfs2_simple_size_update(struct inode *inode,
 310                              struct buffer_head *di_bh,
 311                              u64 new_i_size)
 312 {
 313         int ret;
 314         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 315         handle_t *handle = NULL;
 316
 317         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
 318         if (IS_ERR(handle)) {
 319                 ret = PTR_ERR(handle);
 320                 mlog_errno(ret);
 321                 goto out;
 322         }
 323
 324         ret = ocfs2_set_inode_size(handle, inode, di_bh,
 325                                    new_i_size);
 326         if (ret < 0)
 327                 mlog_errno(ret);
 328
 329         ocfs2_commit_trans(osb, handle);
 330 out:
 331         return ret;
 332 }
 333
 334 static int ocfs2_orphan_for_truncate(struct ocfs2_super *osb,
 335                                      struct inode *inode,
 336                                      struct buffer_head *fe_bh,
 337                                      u64 new_i_size)
 338 {
 339         int status;
 340         handle_t *handle;
 341         struct ocfs2_dinode *di;
 342         u64 cluster_bytes;
 343
 344         mlog_entry_void();
 345
 346         /* TODO: This needs to actually orphan the inode in this
 347          * transaction. */
 348
 349         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
 350         if (IS_ERR(handle)) {
 351                 status = PTR_ERR(handle);
 352                 mlog_errno(status);
 353                 goto out;
 354         }
 355
 356         status = ocfs2_journal_access_di(handle, inode, fe_bh,
 357                                          OCFS2_JOURNAL_ACCESS_WRITE);
 358         if (status < 0) {
 359                 mlog_errno(status);
 360                 goto out_commit;
 361         }
 362
 363         /*
 364          * Do this before setting i_size.
 365          */
 366         cluster_bytes = ocfs2_align_bytes_to_clusters(inode->i_sb, new_i_size);
 367         status = ocfs2_zero_range_for_truncate(inode, handle, new_i_size,
 368                                                cluster_bytes);
 369         if (status) {
 370                 mlog_errno(status);
 371                 goto out_commit;
 372         }
 373
 374         i_size_write(inode, new_i_size);
 375         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
 376
 377         di = (struct ocfs2_dinode *) fe_bh->b_data;
 378         di->i_size = cpu_to_le64(new_i_size);
 379         di->i_ctime = di->i_mtime = cpu_to_le64(inode->i_ctime.tv_sec);
 380         di->i_ctime_nsec = di->i_mtime_nsec = cpu_to_le32(inode->i_ctime.tv_nsec);
 381
 382         status = ocfs2_journal_dirty(handle, fe_bh);
 383         if (status < 0)
 384                 mlog_errno(status);
 385
 386 out_commit:
 387         ocfs2_commit_trans(osb, handle);
 388 out:
 389
 390         mlog_exit(status);
 391         return status;
 392 }
 393
 394 static int ocfs2_truncate_file(struct inode *inode,
 395                                struct buffer_head *di_bh,
 396                                u64 new_i_size)
 397 {
 398         int status = 0;
 399         struct ocfs2_dinode *fe = NULL;
 400         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 401         struct ocfs2_truncate_context *tc = NULL;
 402
 403         mlog_entry("(inode = %llu, new_i_size = %llu\n",
 404                    (unsigned long long)OCFS2_I(inode)->ip_blkno,
 405                    (unsigned long long)new_i_size);
 406
 407         /* We trust di_bh because it comes from ocfs2_inode_lock(), which
 408          * already validated it */
 409         fe = (struct ocfs2_dinode *) di_bh->b_data;
 410
 411         mlog_bug_on_msg(le64_to_cpu(fe->i_size) != i_size_read(inode),
 412                         "Inode %llu, inode i_size = %lld != di "
 413                         "i_size = %llu, i_flags = 0x%x\n",
 414                         (unsigned long long)OCFS2_I(inode)->ip_blkno,
 415                         i_size_read(inode),
 416                         (unsigned long long)le64_to_cpu(fe->i_size),
 417                         le32_to_cpu(fe->i_flags));
 418
 419         if (new_i_size > le64_to_cpu(fe->i_size)) {
 420                 mlog(0, "asked to truncate file with size (%llu) to size (%llu)!\n",
 421                      (unsigned long long)le64_to_cpu(fe->i_size),
 422                      (unsigned long long)new_i_size);
 423                 status = -EINVAL;
 424                 mlog_errno(status);
 425                 goto bail;
 426         }
 427
 428         mlog(0, "inode %llu, i_size = %llu, new_i_size = %llu\n",
 429              (unsigned long long)le64_to_cpu(fe->i_blkno),
 430              (unsigned long long)le64_to_cpu(fe->i_size),
 431              (unsigned long long)new_i_size);
 432
 433         /* lets handle the simple truncate cases before doing any more
 434          * cluster locking. */
 435         if (new_i_size == le64_to_cpu(fe->i_size))
 436                 goto bail;
 437
 438         down_write(&OCFS2_I(inode)->ip_alloc_sem);
 439
 440         /*
 441          * The inode lock forced other nodes to sync and drop their
 442          * pages, which (correctly) happens even if we have a truncate
 443          * without allocation change - ocfs2 cluster sizes can be much
 444          * greater than page size, so we have to truncate them
 445          * anyway.
 446          */
 447         unmap_mapping_range(inode->i_mapping, new_i_size + PAGE_SIZE - 1, 0, 1);
 448         truncate_inode_pages(inode->i_mapping, new_i_size);
 449
 450         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
 451                 status = ocfs2_truncate_inline(inode, di_bh, new_i_size,
 452                                                i_size_read(inode), 1);
 453                 if (status)
 454                         mlog_errno(status);
 455
 456                 goto bail_unlock_sem;
 457         }
 458
 459         /* alright, we're going to need to do a full blown alloc size
 460          * change. Orphan the inode so that recovery can complete the
 461          * truncate if necessary. This does the task of marking
 462          * i_size. */
 463         status = ocfs2_orphan_for_truncate(osb, inode, di_bh, new_i_size);
 464         if (status < 0) {
 465                 mlog_errno(status);
 466                 goto bail_unlock_sem;
 467         }
 468
 469         status = ocfs2_prepare_truncate(osb, inode, di_bh, &tc);
 470         if (status < 0) {
 471                 mlog_errno(status);
 472                 goto bail_unlock_sem;
 473         }
 474
 475         status = ocfs2_commit_truncate(osb, inode, di_bh, tc);
 476         if (status < 0) {
 477                 mlog_errno(status);
 478                 goto bail_unlock_sem;
 479         }
 480
 481         /* TODO: orphan dir cleanup here. */
 482 bail_unlock_sem:
 483         up_write(&OCFS2_I(inode)->ip_alloc_sem);
 484
 485 bail:
 486
 487         mlog_exit(status);
 488         return status;
 489 }
 490
 491 /*
 492  * extend file allocation only here.
 493  * we'll update all the disk stuff, and oip->alloc_size
 494  *
 495  * expect stuff to be locked, a transaction started and enough data /
 496  * metadata reservations in the contexts.
 497  *
 498  * Will return -EAGAIN, and a reason if a restart is needed.
 499  * If passed in, *reason will always be set, even in error.
 500  */
 501 int ocfs2_add_inode_data(struct ocfs2_super *osb,
 502                          struct inode *inode,
 503                          u32 *logical_offset,
 504                          u32 clusters_to_add,
 505                          int mark_unwritten,
 506                          struct buffer_head *fe_bh,
 507                          handle_t *handle,
 508                          struct ocfs2_alloc_context *data_ac,
 509                          struct ocfs2_alloc_context *meta_ac,
 510                          enum ocfs2_alloc_restarted *reason_ret)
 511 {
 512         int ret;
 513         struct ocfs2_extent_tree et;
 514
 515         ocfs2_init_dinode_extent_tree(&et, inode, fe_bh);
 516         ret = ocfs2_add_clusters_in_btree(osb, inode, logical_offset,
 517                                            clusters_to_add, mark_unwritten,
 518                                            &et, handle,
 519                                            data_ac, meta_ac, reason_ret);
 520
 521         return ret;
 522 }
 523
 524 static int __ocfs2_extend_allocation(struct inode *inode, u32 logical_start,
 525                                      u32 clusters_to_add, int mark_unwritten)
 526 {
 527         int status = 0;
 528         int restart_func = 0;
 529         int credits;
 530         u32 prev_clusters;
 531         struct buffer_head *bh = NULL;
 532         struct ocfs2_dinode *fe = NULL;
 533         handle_t *handle = NULL;
 534         struct ocfs2_alloc_context *data_ac = NULL;
 535         struct ocfs2_alloc_context *meta_ac = NULL;
 536         enum ocfs2_alloc_restarted why;
 537         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 538         struct ocfs2_extent_tree et;
 539         int did_quota = 0;
 540
 541         mlog_entry("(clusters_to_add = %u)\n", clusters_to_add);
 542
 543         /*
 544          * This function only exists for file systems which don't
 545          * support holes.
 546          */
 547         BUG_ON(mark_unwritten && !ocfs2_sparse_alloc(osb));
 548
 549         status = ocfs2_read_inode_block(inode, &bh);
 550         if (status < 0) {
 551                 mlog_errno(status);
 552                 goto leave;
 553         }
 554         fe = (struct ocfs2_dinode *) bh->b_data;
 555
 556 restart_all:
 557         BUG_ON(le32_to_cpu(fe->i_clusters) != OCFS2_I(inode)->ip_clusters);
 558
 559         mlog(0, "extend inode %llu, i_size = %lld, di->i_clusters = %u, "
 560              "clusters_to_add = %u\n",
 561              (unsigned long long)OCFS2_I(inode)->ip_blkno,
 562              (long long)i_size_read(inode), le32_to_cpu(fe->i_clusters),
 563              clusters_to_add);
 564         ocfs2_init_dinode_extent_tree(&et, inode, bh);
 565         status = ocfs2_lock_allocators(inode, &et, clusters_to_add, 0,
 566                                        &data_ac, &meta_ac);
 567         if (status) {
 568                 mlog_errno(status);
 569                 goto leave;
 570         }
 571
 572         credits = ocfs2_calc_extend_credits(osb->sb, &fe->id2.i_list,
 573                                             clusters_to_add);
 574         handle = ocfs2_start_trans(osb, credits);
 575         if (IS_ERR(handle)) {
 576                 status = PTR_ERR(handle);
 577                 handle = NULL;
 578                 mlog_errno(status);
 579                 goto leave;
 580         }
 581
 582 restarted_transaction:
 583         if (vfs_dq_alloc_space_nodirty(inode, ocfs2_clusters_to_bytes(osb->sb,
 584             clusters_to_add))) {
 585                 status = -EDQUOT;
 586                 goto leave;
 587         }
 588         did_quota = 1;
 589
 590         /* reserve a write to the file entry early on - that we if we
 591          * run out of credits in the allocation path, we can still
 592          * update i_size. */
 593         status = ocfs2_journal_access_di(handle, inode, bh,
 594                                          OCFS2_JOURNAL_ACCESS_WRITE);
 595         if (status < 0) {
 596                 mlog_errno(status);
 597                 goto leave;
 598         }
 599
 600         prev_clusters = OCFS2_I(inode)->ip_clusters;
 601
 602         status = ocfs2_add_inode_data(osb,
 603                                       inode,
 604                                       &logical_start,
 605                                       clusters_to_add,
 606                                       mark_unwritten,
 607                                       bh,
 608                                       handle,
 609                                       data_ac,
 610                                       meta_ac,
 611                                       &why);
 612         if ((status < 0) && (status != -EAGAIN)) {
 613                 if (status != -ENOSPC)
 614                         mlog_errno(status);
 615                 goto leave;
 616         }
 617
 618         status = ocfs2_journal_dirty(handle, bh);
 619         if (status < 0) {
 620                 mlog_errno(status);
 621                 goto leave;
 622         }
 623
 624         spin_lock(&OCFS2_I(inode)->ip_lock);
 625         clusters_to_add -= (OCFS2_I(inode)->ip_clusters - prev_clusters);
 626         spin_unlock(&OCFS2_I(inode)->ip_lock);
 627         /* Release unused quota reservation */
 628         vfs_dq_free_space(inode,
 629                         ocfs2_clusters_to_bytes(osb->sb, clusters_to_add));
 630         did_quota = 0;
 631
 632         if (why != RESTART_NONE && clusters_to_add) {
 633                 if (why == RESTART_META) {
 634                         mlog(0, "restarting function.\n");
 635                         restart_func = 1;
 636                 } else {
 637                         BUG_ON(why != RESTART_TRANS);
 638
 639                         mlog(0, "restarting transaction.\n");
 640                         /* TODO: This can be more intelligent. */
 641                         credits = ocfs2_calc_extend_credits(osb->sb,
 642                                                             &fe->id2.i_list,
 643                                                             clusters_to_add);
 644                         status = ocfs2_extend_trans(handle, credits);
 645                         if (status < 0) {
 646                                 /* handle still has to be committed at
 647                                  * this point. */
 648                                 status = -ENOMEM;
 649                                 mlog_errno(status);
 650                                 goto leave;
 651                         }
 652                         goto restarted_transaction;
 653                 }
 654         }
 655
 656         mlog(0, "fe: i_clusters = %u, i_size=%llu\n",
 657              le32_to_cpu(fe->i_clusters),
 658              (unsigned long long)le64_to_cpu(fe->i_size));
 659         mlog(0, "inode: ip_clusters=%u, i_size=%lld\n",
 660              OCFS2_I(inode)->ip_clusters, (long long)i_size_read(inode));
 661
 662 leave:
 663         if (status < 0 && did_quota)
 664                 vfs_dq_free_space(inode,
 665                         ocfs2_clusters_to_bytes(osb->sb, clusters_to_add));
 666         if (handle) {
 667                 ocfs2_commit_trans(osb, handle);
 668                 handle = NULL;
 669         }
 670         if (data_ac) {
 671                 ocfs2_free_alloc_context(data_ac);
 672                 data_ac = NULL;
 673         }
 674         if (meta_ac) {
 675                 ocfs2_free_alloc_context(meta_ac);
 676                 meta_ac = NULL;
 677         }
 678         if ((!status) && restart_func) {
 679                 restart_func = 0;
 680                 goto restart_all;
 681         }
 682         brelse(bh);
 683         bh = NULL;
 684
 685         mlog_exit(status);
 686         return status;
 687 }
 688
 689 /* Some parts of this taken from generic_cont_expand, which turned out
 690  * to be too fragile to do exactly what we need without us having to
 691  * worry about recursive locking in ->write_begin() and ->write_end(). */
 692 static int ocfs2_write_zero_page(struct inode *inode,
 693                                  u64 size)
 694 {
 695         struct address_space *mapping = inode->i_mapping;
 696         struct page *page;
 697         unsigned long index;
 698         unsigned int offset;
 699         handle_t *handle = NULL;
 700         int ret;
 701
 702         offset = (size & (PAGE_CACHE_SIZE-1)); /* Within page */
 703         /* ugh.  in prepare/commit_write, if from==to==start of block, we
 704         ** skip the prepare.  make sure we never send an offset for the start
 705         ** of a block
 706         */
 707         if ((offset & (inode->i_sb->s_blocksize - 1)) == 0) {
 708                 offset++;
 709         }
 710         index = size >> PAGE_CACHE_SHIFT;
 711
 712         page = grab_cache_page(mapping, index);
 713         if (!page) {
 714                 ret = -ENOMEM;
 715                 mlog_errno(ret);
 716                 goto out;
 717         }
 718
 719         ret = ocfs2_prepare_write_nolock(inode, page, offset, offset);
 720         if (ret < 0) {
 721                 mlog_errno(ret);
 722                 goto out_unlock;
 723         }
 724
 725         if (ocfs2_should_order_data(inode)) {
 726                 handle = ocfs2_start_walk_page_trans(inode, page, offset,
 727                                                      offset);
 728                 if (IS_ERR(handle)) {
 729                         ret = PTR_ERR(handle);
 730                         handle = NULL;
 731                         goto out_unlock;
 732                 }
 733         }
 734
 735         /* must not update i_size! */
 736         ret = block_commit_write(page, offset, offset);
 737         if (ret < 0)
 738                 mlog_errno(ret);
 739         else
 740                 ret = 0;
 741
 742         if (handle)
 743                 ocfs2_commit_trans(OCFS2_SB(inode->i_sb), handle);
 744 out_unlock:
 745         unlock_page(page);
 746         page_cache_release(page);
 747 out:
 748         return ret;
 749 }
 750
 751 static int ocfs2_zero_extend(struct inode *inode,
 752                              u64 zero_to_size)
 753 {
 754         int ret = 0;
 755         u64 start_off;
 756         struct super_block *sb = inode->i_sb;
 757
 758         start_off = ocfs2_align_bytes_to_blocks(sb, i_size_read(inode));
 759         while (start_off < zero_to_size) {
 760                 ret = ocfs2_write_zero_page(inode, start_off);
 761                 if (ret < 0) {
 762                         mlog_errno(ret);
 763                         goto out;
 764                 }
 765
 766                 start_off += sb->s_blocksize;
 767
 768                 /*
 769                  * Very large extends have the potential to lock up
 770                  * the cpu for extended periods of time.
 771                  */
 772                 cond_resched();
 773         }
 774
 775 out:
 776         return ret;
 777 }
 778
 779 int ocfs2_extend_no_holes(struct inode *inode, u64 new_i_size, u64 zero_to)
 780 {
 781         int ret;
 782         u32 clusters_to_add;
 783         struct ocfs2_inode_info *oi = OCFS2_I(inode);
 784
 785         clusters_to_add = ocfs2_clusters_for_bytes(inode->i_sb, new_i_size);
 786         if (clusters_to_add < oi->ip_clusters)
 787                 clusters_to_add = 0;
 788         else
 789                 clusters_to_add -= oi->ip_clusters;
 790
 791         if (clusters_to_add) {
 792                 ret = __ocfs2_extend_allocation(inode, oi->ip_clusters,
 793                                                 clusters_to_add, 0);
 794                 if (ret) {
 795                         mlog_errno(ret);
 796                         goto out;
 797                 }
 798         }
 799
 800         /*
 801          * Call this even if we don't add any clusters to the tree. We
 802          * still need to zero the area between the old i_size and the
 803          * new i_size.
 804          */
 805         ret = ocfs2_zero_extend(inode, zero_to);
 806         if (ret < 0)
 807                 mlog_errno(ret);
 808
 809 out:
 810         return ret;
 811 }
 812
 813 static int ocfs2_extend_file(struct inode *inode,
 814                              struct buffer_head *di_bh,
 815                              u64 new_i_size)
 816 {
 817         int ret = 0;
 818         struct ocfs2_inode_info *oi = OCFS2_I(inode);
 819
 820         BUG_ON(!di_bh);
 821
 822         /* setattr sometimes calls us like this. */
 823         if (new_i_size == 0)
 824                 goto out;
 825
 826         if (i_size_read(inode) == new_i_size)
 827                 goto out;
 828         BUG_ON(new_i_size < i_size_read(inode));
 829
 830         /*
 831          * Fall through for converting inline data, even if the fs
 832          * supports sparse files.
 833          *
 834          * The check for inline data here is legal - nobody can add
 835          * the feature since we have i_mutex. We must check it again
 836          * after acquiring ip_alloc_sem though, as paths like mmap
 837          * might have raced us to converting the inode to extents.
 838          */
 839         if (!(oi->ip_dyn_features & OCFS2_INLINE_DATA_FL)
 840             && ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)))
 841                 goto out_update_size;
 842
 843         /*
 844          * The alloc sem blocks people in read/write from reading our
 845          * allocation until we're done changing it. We depend on
 846          * i_mutex to block other extend/truncate calls while we're
 847          * here.
 848          */
 849         down_write(&oi->ip_alloc_sem);
 850
 851         if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
 852                 /*
 853                  * We can optimize small extends by keeping the inodes
 854                  * inline data.
 855                  */
 856                 if (ocfs2_size_fits_inline_data(di_bh, new_i_size)) {
 857                         up_write(&oi->ip_alloc_sem);
 858                         goto out_update_size;
 859                 }
 860
 861                 ret = ocfs2_convert_inline_data_to_extents(inode, di_bh);
 862                 if (ret) {
 863                         up_write(&oi->ip_alloc_sem);
 864
 865                         mlog_errno(ret);
 866                         goto out;
 867                 }
 868         }
 869
 870         if (!ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)))
 871                 ret = ocfs2_extend_no_holes(inode, new_i_size, new_i_size);
 872
 873         up_write(&oi->ip_alloc_sem);
 874
 875         if (ret < 0) {
 876                 mlog_errno(ret);
 877                 goto out;
 878         }
 879
 880 out_update_size:
 881         ret = ocfs2_simple_size_update(inode, di_bh, new_i_size);
 882         if (ret < 0)
 883                 mlog_errno(ret);
 884
 885 out:
 886         return ret;
 887 }
 888
 889 int ocfs2_setattr(struct dentry *dentry, struct iattr *attr)
 890 {
 891         int status = 0, size_change;
 892         struct inode *inode = dentry->d_inode;
 893         struct super_block *sb = inode->i_sb;
 894         struct ocfs2_super *osb = OCFS2_SB(sb);
 895         struct buffer_head *bh = NULL;
 896         handle_t *handle = NULL;
 897         int locked[MAXQUOTAS] = {0, 0};
 898         int credits, qtype;
 899         struct ocfs2_mem_dqinfo *oinfo;
 900
 901         mlog_entry("(0x%p, '%.*s')\n", dentry,
 902                    dentry->d_name.len, dentry->d_name.name);
 903
 904         /* ensuring we don't even attempt to truncate a symlink */
 905         if (S_ISLNK(inode->i_mode))
 906                 attr->ia_valid &= ~ATTR_SIZE;
 907
 908         if (attr->ia_valid & ATTR_MODE)
 909                 mlog(0, "mode change: %d\n", attr->ia_mode);
 910         if (attr->ia_valid & ATTR_UID)
 911                 mlog(0, "uid change: %d\n", attr->ia_uid);
 912         if (attr->ia_valid & ATTR_GID)
 913                 mlog(0, "gid change: %d\n", attr->ia_gid);
 914         if (attr->ia_valid & ATTR_SIZE)
 915                 mlog(0, "size change...\n");
 916         if (attr->ia_valid & (ATTR_ATIME | ATTR_MTIME | ATTR_CTIME))
 917                 mlog(0, "time change...\n");
 918
 919 #define OCFS2_VALID_ATTRS (ATTR_ATIME | ATTR_MTIME | ATTR_CTIME | ATTR_SIZE \
 920                            | ATTR_GID | ATTR_UID | ATTR_MODE)
 921         if (!(attr->ia_valid & OCFS2_VALID_ATTRS)) {
 922                 mlog(0, "can't handle attrs: 0x%x\n", attr->ia_valid);
 923                 return 0;
 924         }
 925
 926         status = inode_change_ok(inode, attr);
 927         if (status)
 928                 return status;
 929
 930         size_change = S_ISREG(inode->i_mode) && attr->ia_valid & ATTR_SIZE;
 931         if (size_change) {
 932                 status = ocfs2_rw_lock(inode, 1);
 933                 if (status < 0) {
 934                         mlog_errno(status);
 935                         goto bail;
 936                 }
 937         }
 938
 939         status = ocfs2_inode_lock(inode, &bh, 1);
 940         if (status < 0) {
 941                 if (status != -ENOENT)
 942                         mlog_errno(status);
 943                 goto bail_unlock_rw;
 944         }
 945
 946         if (size_change && attr->ia_size != i_size_read(inode)) {
 947                 if (attr->ia_size > sb->s_maxbytes) {
 948                         status = -EFBIG;
 949                         goto bail_unlock;
 950                 }
 951
 952                 if (i_size_read(inode) > attr->ia_size) {
 953                         if (ocfs2_should_order_data(inode)) {
 954                                 status = ocfs2_begin_ordered_truncate(inode,
 955                                                                       attr->ia_size);
 956                                 if (status)
 957                                         goto bail_unlock;
 958                         }
 959                         status = ocfs2_truncate_file(inode, bh, attr->ia_size);
 960                 } else
 961                         status = ocfs2_extend_file(inode, bh, attr->ia_size);
 962                 if (status < 0) {
 963                         if (status != -ENOSPC)
 964                                 mlog_errno(status);
 965                         status = -ENOSPC;
 966                         goto bail_unlock;
 967                 }
 968         }
 969
 970         if ((attr->ia_valid & ATTR_UID && attr->ia_uid != inode->i_uid) ||
 971             (attr->ia_valid & ATTR_GID && attr->ia_gid != inode->i_gid)) {
 972                 credits = OCFS2_INODE_UPDATE_CREDITS;
 973                 if (attr->ia_valid & ATTR_UID && attr->ia_uid != inode->i_uid
 974                     && OCFS2_HAS_RO_COMPAT_FEATURE(sb,
 975                     OCFS2_FEATURE_RO_COMPAT_USRQUOTA)) {
 976                         oinfo = sb_dqinfo(sb, USRQUOTA)->dqi_priv;
 977                         status = ocfs2_lock_global_qf(oinfo, 1);
 978                         if (status < 0)
 979                                 goto bail_unlock;
 980                         credits += ocfs2_calc_qinit_credits(sb, USRQUOTA) +
 981                                 ocfs2_calc_qdel_credits(sb, USRQUOTA);
 982                         locked[USRQUOTA] = 1;
 983                 }
 984                 if (attr->ia_valid & ATTR_GID && attr->ia_gid != inode->i_gid
 985                     && OCFS2_HAS_RO_COMPAT_FEATURE(sb,
 986                     OCFS2_FEATURE_RO_COMPAT_GRPQUOTA)) {
 987                         oinfo = sb_dqinfo(sb, GRPQUOTA)->dqi_priv;
 988                         status = ocfs2_lock_global_qf(oinfo, 1);
 989                         if (status < 0)
 990                                 goto bail_unlock;
 991                         credits += ocfs2_calc_qinit_credits(sb, GRPQUOTA) +
 992                                    ocfs2_calc_qdel_credits(sb, GRPQUOTA);
 993                         locked[GRPQUOTA] = 1;
 994                 }
 995                 handle = ocfs2_start_trans(osb, credits);
 996                 if (IS_ERR(handle)) {
 997                         status = PTR_ERR(handle);
 998                         mlog_errno(status);
 999                         goto bail_unlock;
1000                 }
1001                 status = vfs_dq_transfer(inode, attr) ? -EDQUOT : 0;
1002                 if (status < 0)
1003                         goto bail_commit;
1004         } else {
1005                 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1006                 if (IS_ERR(handle)) {
1007                         status = PTR_ERR(handle);
1008                         mlog_errno(status);
1009                         goto bail_unlock;
1010                 }
1011         }
1012
1013         /*
1014          * This will intentionally not wind up calling vmtruncate(),
1015          * since all the work for a size change has been done above.
1016          * Otherwise, we could get into problems with truncate as
1017          * ip_alloc_sem is used there to protect against i_size
1018          * changes.
1019          */
1020         status = inode_setattr(inode, attr);
1021         if (status < 0) {
1022                 mlog_errno(status);
1023                 goto bail_commit;
1024         }
1025
1026         status = ocfs2_mark_inode_dirty(handle, inode, bh);
1027         if (status < 0)
1028                 mlog_errno(status);
1029
1030 bail_commit:
1031         ocfs2_commit_trans(osb, handle);
1032 bail_unlock:
1033         for (qtype = 0; qtype < MAXQUOTAS; qtype++) {
1034                 if (!locked[qtype])
1035                         continue;
1036                 oinfo = sb_dqinfo(sb, qtype)->dqi_priv;
1037                 ocfs2_unlock_global_qf(oinfo, 1);
1038         }
1039         ocfs2_inode_unlock(inode, 1);
1040 bail_unlock_rw:
1041         if (size_change)
1042                 ocfs2_rw_unlock(inode, 1);
1043 bail:
1044         brelse(bh);
1045
1046         if (!status && attr->ia_valid & ATTR_MODE) {
1047                 status = ocfs2_acl_chmod(inode);
1048                 if (status < 0)
1049                         mlog_errno(status);
1050         }
1051
1052         mlog_exit(status);
1053         return status;
1054 }
1055
1056 int ocfs2_getattr(struct vfsmount *mnt,
1057                   struct dentry *dentry,
1058                   struct kstat *stat)
1059 {
1060         struct inode *inode = dentry->d_inode;
1061         struct super_block *sb = dentry->d_inode->i_sb;
1062         struct ocfs2_super *osb = sb->s_fs_info;
1063         int err;
1064
1065         mlog_entry_void();
1066
1067         err = ocfs2_inode_revalidate(dentry);
1068         if (err) {
1069                 if (err != -ENOENT)
1070                         mlog_errno(err);
1071                 goto bail;
1072         }
1073
1074         generic_fillattr(inode, stat);
1075
1076         /* We set the blksize from the cluster size for performance */
1077         stat->blksize = osb->s_clustersize;
1078
1079 bail:
1080         mlog_exit(err);
1081
1082         return err;
1083 }
1084
1085 int ocfs2_permission(struct inode *inode, int mask)
1086 {
1087         int ret;
1088
1089         mlog_entry_void();
1090
1091         ret = ocfs2_inode_lock(inode, NULL, 0);
1092         if (ret) {
1093                 if (ret != -ENOENT)
1094                         mlog_errno(ret);
1095                 goto out;
1096         }
1097
1098         ret = generic_permission(inode, mask, ocfs2_check_acl);
1099
1100         ocfs2_inode_unlock(inode, 0);
1101 out:
1102         mlog_exit(ret);
1103         return ret;
1104 }
1105
1106 static int __ocfs2_write_remove_suid(struct inode *inode,
1107                                      struct buffer_head *bh)
1108 {
1109         int ret;
1110         handle_t *handle;
1111         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1112         struct ocfs2_dinode *di;
1113
1114         mlog_entry("(Inode %llu, mode 0%o)\n",
1115                    (unsigned long long)OCFS2_I(inode)->ip_blkno, inode->i_mode);
1116
1117         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1118         if (IS_ERR(handle)) {
1119                 ret = PTR_ERR(handle);
1120                 mlog_errno(ret);
1121                 goto out;
1122         }
1123
1124         ret = ocfs2_journal_access_di(handle, inode, bh,
1125                                       OCFS2_JOURNAL_ACCESS_WRITE);
1126         if (ret < 0) {
1127                 mlog_errno(ret);
1128                 goto out_trans;
1129         }
1130
1131         inode->i_mode &= ~S_ISUID;
1132         if ((inode->i_mode & S_ISGID) && (inode->i_mode & S_IXGRP))
1133                 inode->i_mode &= ~S_ISGID;
1134
1135         di = (struct ocfs2_dinode *) bh->b_data;
1136         di->i_mode = cpu_to_le16(inode->i_mode);
1137
1138         ret = ocfs2_journal_dirty(handle, bh);
1139         if (ret < 0)
1140                 mlog_errno(ret);
1141
1142 out_trans:
1143         ocfs2_commit_trans(osb, handle);
1144 out:
1145         mlog_exit(ret);
1146         return ret;
1147 }
1148
1149 /*
1150  * Will look for holes and unwritten extents in the range starting at
1151  * pos for count bytes (inclusive).
1152  */
1153 static int ocfs2_check_range_for_holes(struct inode *inode, loff_t pos,
1154                                        size_t count)
1155 {
1156         int ret = 0;
1157         unsigned int extent_flags;
1158         u32 cpos, clusters, extent_len, phys_cpos;
1159         struct super_block *sb = inode->i_sb;
1160
1161         cpos = pos >> OCFS2_SB(sb)->s_clustersize_bits;
1162         clusters = ocfs2_clusters_for_bytes(sb, pos + count) - cpos;
1163
1164         while (clusters) {
1165                 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos, &extent_len,
1166                                          &extent_flags);
1167                 if (ret < 0) {
1168                         mlog_errno(ret);
1169                         goto out;
1170                 }
1171
1172                 if (phys_cpos == 0 || (extent_flags & OCFS2_EXT_UNWRITTEN)) {
1173                         ret = 1;
1174                         break;
1175                 }
1176
1177                 if (extent_len > clusters)
1178                         extent_len = clusters;
1179
1180                 clusters -= extent_len;
1181                 cpos += extent_len;
1182         }
1183 out:
1184         return ret;
1185 }
1186
1187 static int ocfs2_write_remove_suid(struct inode *inode)
1188 {
1189         int ret;
1190         struct buffer_head *bh = NULL;
1191
1192         ret = ocfs2_read_inode_block(inode, &bh);
1193         if (ret < 0) {
1194                 mlog_errno(ret);
1195                 goto out;
1196         }
1197
1198         ret =  __ocfs2_write_remove_suid(inode, bh);
1199 out:
1200         brelse(bh);
1201         return ret;
1202 }
1203
1204 /*
1205  * Allocate enough extents to cover the region starting at byte offset
1206  * start for len bytes. Existing extents are skipped, any extents
1207  * added are marked as "unwritten".
1208  */
1209 static int ocfs2_allocate_unwritten_extents(struct inode *inode,
1210                                             u64 start, u64 len)
1211 {
1212         int ret;
1213         u32 cpos, phys_cpos, clusters, alloc_size;
1214         u64 end = start + len;
1215         struct buffer_head *di_bh = NULL;
1216
1217         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1218                 ret = ocfs2_read_inode_block(inode, &di_bh);
1219                 if (ret) {
1220                         mlog_errno(ret);
1221                         goto out;
1222                 }
1223
1224                 /*
1225                  * Nothing to do if the requested reservation range
1226                  * fits within the inode.
1227                  */
1228                 if (ocfs2_size_fits_inline_data(di_bh, end))
1229                         goto out;
1230
1231                 ret = ocfs2_convert_inline_data_to_extents(inode, di_bh);
1232                 if (ret) {
1233                         mlog_errno(ret);
1234                         goto out;
1235                 }
1236         }
1237
1238         /*
1239          * We consider both start and len to be inclusive.
1240          */
1241         cpos = start >> OCFS2_SB(inode->i_sb)->s_clustersize_bits;
1242         clusters = ocfs2_clusters_for_bytes(inode->i_sb, start + len);
1243         clusters -= cpos;
1244
1245         while (clusters) {
1246                 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos,
1247                                          &alloc_size, NULL);
1248                 if (ret) {
1249                         mlog_errno(ret);
1250                         goto out;
1251                 }
1252
1253                 /*
1254                  * Hole or existing extent len can be arbitrary, so
1255                  * cap it to our own allocation request.
1256                  */
1257                 if (alloc_size > clusters)
1258                         alloc_size = clusters;
1259
1260                 if (phys_cpos) {
1261                         /*
1262                          * We already have an allocation at this
1263                          * region so we can safely skip it.
1264                          */
1265                         goto next;
1266                 }
1267
1268                 ret = __ocfs2_extend_allocation(inode, cpos, alloc_size, 1);
1269                 if (ret) {
1270                         if (ret != -ENOSPC)
1271                                 mlog_errno(ret);
1272                         goto out;
1273                 }
1274
1275 next:
1276                 cpos += alloc_size;
1277                 clusters -= alloc_size;
1278         }
1279
1280         ret = 0;
1281 out:
1282
1283         brelse(di_bh);
1284         return ret;
1285 }
1286
1287 /*
1288  * Truncate a byte range, avoiding pages within partial clusters. This
1289  * preserves those pages for the zeroing code to write to.
1290  */
1291 static void ocfs2_truncate_cluster_pages(struct inode *inode, u64 byte_start,
1292                                          u64 byte_len)
1293 {
1294         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1295         loff_t start, end;
1296         struct address_space *mapping = inode->i_mapping;
1297
1298         start = (loff_t)ocfs2_align_bytes_to_clusters(inode->i_sb, byte_start);
1299         end = byte_start + byte_len;
1300         end = end & ~(osb->s_clustersize - 1);
1301
1302         if (start < end) {
1303                 unmap_mapping_range(mapping, start, end - start, 0);
1304                 truncate_inode_pages_range(mapping, start, end - 1);
1305         }
1306 }
1307
1308 static int ocfs2_zero_partial_clusters(struct inode *inode,
1309                                        u64 start, u64 len)
1310 {
1311         int ret = 0;
1312         u64 tmpend, end = start + len;
1313         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1314         unsigned int csize = osb->s_clustersize;
1315         handle_t *handle;
1316
1317         /*
1318          * The "start" and "end" values are NOT necessarily part of
1319          * the range whose allocation is being deleted. Rather, this
1320          * is what the user passed in with the request. We must zero
1321          * partial clusters here. There's no need to worry about
1322          * physical allocation - the zeroing code knows to skip holes.
1323          */
1324         mlog(0, "byte start: %llu, end: %llu\n",
1325              (unsigned long long)start, (unsigned long long)end);
1326
1327         /*
1328          * If both edges are on a cluster boundary then there's no
1329          * zeroing required as the region is part of the allocation to
1330          * be truncated.
1331          */
1332         if ((start & (csize - 1)) == 0 && (end & (csize - 1)) == 0)
1333                 goto out;
1334
1335         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1336         if (IS_ERR(handle)) {
1337                 ret = PTR_ERR(handle);
1338                 mlog_errno(ret);
1339                 goto out;
1340         }
1341
1342         /*
1343          * We want to get the byte offset of the end of the 1st cluster.
1344          */
1345         tmpend = (u64)osb->s_clustersize + (start & ~(osb->s_clustersize - 1));
1346         if (tmpend > end)
1347                 tmpend = end;
1348
1349         mlog(0, "1st range: start: %llu, tmpend: %llu\n",
1350              (unsigned long long)start, (unsigned long long)tmpend);
1351
1352         ret = ocfs2_zero_range_for_truncate(inode, handle, start, tmpend);
1353         if (ret)
1354                 mlog_errno(ret);
1355
1356         if (tmpend < end) {
1357                 /*
1358                  * This may make start and end equal, but the zeroing
1359                  * code will skip any work in that case so there's no
1360                  * need to catch it up here.
1361                  */
1362                 start = end & ~(osb->s_clustersize - 1);
1363
1364                 mlog(0, "2nd range: start: %llu, end: %llu\n",
1365                      (unsigned long long)start, (unsigned long long)end);
1366
1367                 ret = ocfs2_zero_range_for_truncate(inode, handle, start, end);
1368                 if (ret)
1369                         mlog_errno(ret);
1370         }
1371
1372         ocfs2_commit_trans(osb, handle);
1373 out:
1374         return ret;
1375 }
1376
1377 static int ocfs2_remove_inode_range(struct inode *inode,
1378                                     struct buffer_head *di_bh, u64 byte_start,
1379                                     u64 byte_len)
1380 {
1381         int ret = 0;
1382         u32 trunc_start, trunc_len, cpos, phys_cpos, alloc_size;
1383         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1384         struct ocfs2_cached_dealloc_ctxt dealloc;
1385         struct address_space *mapping = inode->i_mapping;
1386         struct ocfs2_extent_tree et;
1387
1388         ocfs2_init_dinode_extent_tree(&et, inode, di_bh);
1389         ocfs2_init_dealloc_ctxt(&dealloc);
1390
1391         if (byte_len == 0)
1392                 return 0;
1393
1394         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1395                 ret = ocfs2_truncate_inline(inode, di_bh, byte_start,
1396                                             byte_start + byte_len, 0);
1397                 if (ret) {
1398                         mlog_errno(ret);
1399                         goto out;
1400                 }
1401                 /*
1402                  * There's no need to get fancy with the page cache
1403                  * truncate of an inline-data inode. We're talking
1404                  * about less than a page here, which will be cached
1405                  * in the dinode buffer anyway.
1406                  */
1407                 unmap_mapping_range(mapping, 0, 0, 0);
1408                 truncate_inode_pages(mapping, 0);
1409                 goto out;
1410         }
1411
1412         trunc_start = ocfs2_clusters_for_bytes(osb->sb, byte_start);
1413         trunc_len = (byte_start + byte_len) >> osb->s_clustersize_bits;
1414         if (trunc_len >= trunc_start)
1415                 trunc_len -= trunc_start;
1416         else
1417                 trunc_len = 0;
1418
1419         mlog(0, "Inode: %llu, start: %llu, len: %llu, cstart: %u, clen: %u\n",
1420              (unsigned long long)OCFS2_I(inode)->ip_blkno,
1421              (unsigned long long)byte_start,
1422              (unsigned long long)byte_len, trunc_start, trunc_len);
1423
1424         ret = ocfs2_zero_partial_clusters(inode, byte_start, byte_len);
1425         if (ret) {
1426                 mlog_errno(ret);
1427                 goto out;
1428         }
1429
1430         cpos = trunc_start;
1431         while (trunc_len) {
1432                 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos,
1433                                          &alloc_size, NULL);
1434                 if (ret) {
1435                         mlog_errno(ret);
1436                         goto out;
1437                 }
1438
1439                 if (alloc_size > trunc_len)
1440                         alloc_size = trunc_len;
1441
1442                 /* Only do work for non-holes */
1443                 if (phys_cpos != 0) {
1444                         ret = ocfs2_remove_btree_range(inode, &et, cpos,
1445                                                        phys_cpos, alloc_size,
1446                                                        &dealloc);
1447                         if (ret) {
1448                                 mlog_errno(ret);
1449                                 goto out;
1450                         }
1451                 }
1452
1453                 cpos += alloc_size;
1454                 trunc_len -= alloc_size;
1455         }
1456
1457         ocfs2_truncate_cluster_pages(inode, byte_start, byte_len);
1458
1459 out:
1460         ocfs2_schedule_truncate_log_flush(osb, 1);
1461         ocfs2_run_deallocs(osb, &dealloc);
1462
1463         return ret;
1464 }
1465
1466 /*
1467  * Parts of this function taken from xfs_change_file_space()
1468  */
1469 static int __ocfs2_change_file_space(struct file *file, struct inode *inode,
1470                                      loff_t f_pos, unsigned int cmd,
1471                                      struct ocfs2_space_resv *sr,
1472                                      int change_size)
1473 {
1474         int ret;
1475         s64 llen;
1476         loff_t size;
1477         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1478         struct buffer_head *di_bh = NULL;
1479         handle_t *handle;
1480         unsigned long long max_off = inode->i_sb->s_maxbytes;
1481
1482         if (ocfs2_is_hard_readonly(osb) || ocfs2_is_soft_readonly(osb))
1483                 return -EROFS;
1484
1485         mutex_lock(&inode->i_mutex);
1486
1487         /*
1488          * This prevents concurrent writes on other nodes
1489          */
1490         ret = ocfs2_rw_lock(inode, 1);
1491         if (ret) {
1492                 mlog_errno(ret);
1493                 goto out;
1494         }
1495
1496         ret = ocfs2_inode_lock(inode, &di_bh, 1);
1497         if (ret) {
1498                 mlog_errno(ret);
1499                 goto out_rw_unlock;
1500         }
1501
1502         if (inode->i_flags & (S_IMMUTABLE|S_APPEND)) {
1503                 ret = -EPERM;
1504                 goto out_inode_unlock;
1505         }
1506
1507         switch (sr->l_whence) {
1508         case 0: /*SEEK_SET*/
1509                 break;
1510         case 1: /*SEEK_CUR*/
1511                 sr->l_start += f_pos;
1512                 break;
1513         case 2: /*SEEK_END*/
1514                 sr->l_start += i_size_read(inode);
1515                 break;
1516         default:
1517                 ret = -EINVAL;
1518                 goto out_inode_unlock;
1519         }
1520         sr->l_whence = 0;
1521
1522         llen = sr->l_len > 0 ? sr->l_len - 1 : sr->l_len;
1523
1524         if (sr->l_start < 0
1525             || sr->l_start > max_off
1526             || (sr->l_start + llen) < 0
1527             || (sr->l_start + llen) > max_off) {
1528                 ret = -EINVAL;
1529                 goto out_inode_unlock;
1530         }
1531         size = sr->l_start + sr->l_len;
1532
1533         if (cmd == OCFS2_IOC_RESVSP || cmd == OCFS2_IOC_RESVSP64) {
1534                 if (sr->l_len <= 0) {
1535                         ret = -EINVAL;
1536                         goto out_inode_unlock;
1537                 }
1538         }
1539
1540         if (file && should_remove_suid(file->f_path.dentry)) {
1541                 ret = __ocfs2_write_remove_suid(inode, di_bh);
1542                 if (ret) {
1543                         mlog_errno(ret);
1544                         goto out_inode_unlock;
1545                 }
1546         }
1547
1548         down_write(&OCFS2_I(inode)->ip_alloc_sem);
1549         switch (cmd) {
1550         case OCFS2_IOC_RESVSP:
1551         case OCFS2_IOC_RESVSP64:
1552                 /*
1553                  * This takes unsigned offsets, but the signed ones we
1554                  * pass have been checked against overflow above.
1555                  */
1556                 ret = ocfs2_allocate_unwritten_extents(inode, sr->l_start,
1557                                                        sr->l_len);
1558                 break;
1559         case OCFS2_IOC_UNRESVSP:
1560         case OCFS2_IOC_UNRESVSP64:
1561                 ret = ocfs2_remove_inode_range(inode, di_bh, sr->l_start,
1562                                                sr->l_len);
1563                 break;
1564         default:
1565                 ret = -EINVAL;
1566         }
1567         up_write(&OCFS2_I(inode)->ip_alloc_sem);
1568         if (ret) {
1569                 mlog_errno(ret);
1570                 goto out_inode_unlock;
1571         }
1572
1573         /*
1574          * We update c/mtime for these changes
1575          */
1576         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1577         if (IS_ERR(handle)) {
1578                 ret = PTR_ERR(handle);
1579                 mlog_errno(ret);
1580                 goto out_inode_unlock;
1581         }
1582
1583         if (change_size && i_size_read(inode) < size)
1584                 i_size_write(inode, size);
1585
1586         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
1587         ret = ocfs2_mark_inode_dirty(handle, inode, di_bh);
1588         if (ret < 0)
1589                 mlog_errno(ret);
1590
1591         ocfs2_commit_trans(osb, handle);
1592
1593 out_inode_unlock:
1594         brelse(di_bh);
1595         ocfs2_inode_unlock(inode, 1);
1596 out_rw_unlock:
1597         ocfs2_rw_unlock(inode, 1);
1598
1599 out:
1600         mutex_unlock(&inode->i_mutex);
1601         return ret;
1602 }
1603
1604 int ocfs2_change_file_space(struct file *file, unsigned int cmd,
1605                             struct ocfs2_space_resv *sr)
1606 {
1607         struct inode *inode = file->f_path.dentry->d_inode;
1608         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1609
1610         if ((cmd == OCFS2_IOC_RESVSP || cmd == OCFS2_IOC_RESVSP64) &&
1611             !ocfs2_writes_unwritten_extents(osb))
1612                 return -ENOTTY;
1613         else if ((cmd == OCFS2_IOC_UNRESVSP || cmd == OCFS2_IOC_UNRESVSP64) &&
1614                  !ocfs2_sparse_alloc(osb))
1615                 return -ENOTTY;
1616
1617         if (!S_ISREG(inode->i_mode))
1618                 return -EINVAL;
1619
1620         if (!(file->f_mode & FMODE_WRITE))
1621                 return -EBADF;
1622
1623         return __ocfs2_change_file_space(file, inode, file->f_pos, cmd, sr, 0);
1624 }
1625
1626 static long ocfs2_fallocate(struct inode *inode, int mode, loff_t offset,
1627                             loff_t len)
1628 {
1629         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1630         struct ocfs2_space_resv sr;
1631         int change_size = 1;
1632
1633         if (!ocfs2_writes_unwritten_extents(osb))
1634                 return -EOPNOTSUPP;
1635
1636         if (S_ISDIR(inode->i_mode))
1637                 return -ENODEV;
1638
1639         if (mode & FALLOC_FL_KEEP_SIZE)
1640                 change_size = 0;
1641
1642         sr.l_whence = 0;
1643         sr.l_start = (s64)offset;
1644         sr.l_len = (s64)len;
1645
1646         return __ocfs2_change_file_space(NULL, inode, offset,
1647                                          OCFS2_IOC_RESVSP64, &sr, change_size);
1648 }
1649
1650 static int ocfs2_prepare_inode_for_write(struct dentry *dentry,
1651                                          loff_t *ppos,
1652                                          size_t count,
1653                                          int appending,
1654                                          int *direct_io)
1655 {
1656         int ret = 0, meta_level = 0;
1657         struct inode *inode = dentry->d_inode;
1658         loff_t saved_pos, end;
1659
1660         /*
1661          * We start with a read level meta lock and only jump to an ex
1662          * if we need to make modifications here.
1663          */
1664         for(;;) {
1665                 ret = ocfs2_inode_lock(inode, NULL, meta_level);
1666                 if (ret < 0) {
1667                         meta_level = -1;
1668                         mlog_errno(ret);
1669                         goto out;
1670                 }
1671
1672                 /* Clear suid / sgid if necessary. We do this here
1673                  * instead of later in the write path because
1674                  * remove_suid() calls ->setattr without any hint that
1675                  * we may have already done our cluster locking. Since
1676                  * ocfs2_setattr() *must* take cluster locks to
1677                  * proceeed, this will lead us to recursively lock the
1678                  * inode. There's also the dinode i_size state which
1679                  * can be lost via setattr during extending writes (we
1680                  * set inode->i_size at the end of a write. */
1681                 if (should_remove_suid(dentry)) {
1682                         if (meta_level == 0) {
1683                                 ocfs2_inode_unlock(inode, meta_level);
1684                                 meta_level = 1;
1685                                 continue;
1686                         }
1687
1688                         ret = ocfs2_write_remove_suid(inode);
1689                         if (ret < 0) {
1690                                 mlog_errno(ret);
1691                                 goto out_unlock;
1692                         }
1693                 }
1694
1695                 /* work on a copy of ppos until we're sure that we won't have
1696                  * to recalculate it due to relocking. */
1697                 if (appending) {
1698                         saved_pos = i_size_read(inode);
1699                         mlog(0, "O_APPEND: inode->i_size=%llu\n", saved_pos);
1700                 } else {
1701                         saved_pos = *ppos;
1702                 }
1703
1704                 end = saved_pos + count;
1705
1706                 /*
1707                  * Skip the O_DIRECT checks if we don't need
1708                  * them.
1709                  */
1710                 if (!direct_io || !(*direct_io))
1711                         break;
1712
1713                 /*
1714                  * There's no sane way to do direct writes to an inode
1715                  * with inline data.
1716                  */
1717                 if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1718                         *direct_io = 0;
1719                         break;
1720                 }
1721
1722                 /*
1723                  * Allowing concurrent direct writes means
1724                  * i_size changes wouldn't be synchronized, so
1725                  * one node could wind up truncating another
1726                  * nodes writes.
1727                  */
1728                 if (end > i_size_read(inode)) {
1729                         *direct_io = 0;
1730                         break;
1731                 }
1732
1733                 /*
1734                  * We don't fill holes during direct io, so
1735                  * check for them here. If any are found, the
1736                  * caller will have to retake some cluster
1737                  * locks and initiate the io as buffered.
1738                  */
1739                 ret = ocfs2_check_range_for_holes(inode, saved_pos, count);
1740                 if (ret == 1) {
1741                         *direct_io = 0;
1742                         ret = 0;
1743                 } else if (ret < 0)
1744                         mlog_errno(ret);
1745                 break;
1746         }
1747
1748         if (appending)
1749                 *ppos = saved_pos;
1750
1751 out_unlock:
1752         ocfs2_inode_unlock(inode, meta_level);
1753
1754 out:
1755         return ret;
1756 }
1757
1758 static ssize_t ocfs2_file_aio_write(struct kiocb *iocb,
1759                                     const struct iovec *iov,
1760                                     unsigned long nr_segs,
1761                                     loff_t pos)
1762 {
1763         int ret, direct_io, appending, rw_level, have_alloc_sem  = 0;
1764         int can_do_direct;
1765         ssize_t written = 0;
1766         size_t ocount;          /* original count */
1767         size_t count;           /* after file limit checks */
1768         loff_t old_size, *ppos = &iocb->ki_pos;
1769         u32 old_clusters;
1770         struct file *file = iocb->ki_filp;
1771         struct inode *inode = file->f_path.dentry->d_inode;
1772         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1773
1774         mlog_entry("(0x%p, %u, '%.*s')\n", file,
1775                    (unsigned int)nr_segs,
1776                    file->f_path.dentry->d_name.len,
1777                    file->f_path.dentry->d_name.name);
1778
1779         if (iocb->ki_left == 0)
1780                 return 0;
1781
1782         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
1783
1784         appending = file->f_flags & O_APPEND ? 1 : 0;
1785         direct_io = file->f_flags & O_DIRECT ? 1 : 0;
1786
1787         mutex_lock(&inode->i_mutex);
1788
1789 relock:
1790         /* to match setattr's i_mutex -> i_alloc_sem -> rw_lock ordering */
1791         if (direct_io) {
1792                 down_read(&inode->i_alloc_sem);
1793                 have_alloc_sem = 1;
1794         }
1795
1796         /* concurrent O_DIRECT writes are allowed */
1797         rw_level = !direct_io;
1798         ret = ocfs2_rw_lock(inode, rw_level);
1799         if (ret < 0) {
1800                 mlog_errno(ret);
1801                 goto out_sems;
1802         }
1803
1804         can_do_direct = direct_io;
1805         ret = ocfs2_prepare_inode_for_write(file->f_path.dentry, ppos,
1806                                             iocb->ki_left, appending,
1807                                             &can_do_direct);
1808         if (ret < 0) {
1809                 mlog_errno(ret);
1810                 goto out;
1811         }
1812
1813         /*
1814          * We can't complete the direct I/O as requested, fall back to
1815          * buffered I/O.
1816          */
1817         if (direct_io && !can_do_direct) {
1818                 ocfs2_rw_unlock(inode, rw_level);
1819                 up_read(&inode->i_alloc_sem);
1820
1821                 have_alloc_sem = 0;
1822                 rw_level = -1;
1823
1824                 direct_io = 0;
1825                 goto relock;
1826         }
1827
1828         /*
1829          * To later detect whether a journal commit for sync writes is
1830          * necessary, we sample i_size, and cluster count here.
1831          */
1832         old_size = i_size_read(inode);
1833         old_clusters = OCFS2_I(inode)->ip_clusters;
1834
1835         /* communicate with ocfs2_dio_end_io */
1836         ocfs2_iocb_set_rw_locked(iocb, rw_level);
1837
1838         if (direct_io) {
1839                 ret = generic_segment_checks(iov, &nr_segs, &ocount,
1840                                              VERIFY_READ);
1841                 if (ret)
1842                         goto out_dio;
1843
1844                 ret = generic_write_checks(file, ppos, &count,
1845                                            S_ISBLK(inode->i_mode));
1846                 if (ret)
1847                         goto out_dio;
1848
1849                 written = generic_file_direct_write(iocb, iov, &nr_segs, *ppos,
1850                                                     ppos, count, ocount);
1851                 if (written < 0) {
1852                         /*
1853                          * direct write may have instantiated a few
1854                          * blocks outside i_size. Trim these off again.
1855                          * Don't need i_size_read because we hold i_mutex.
1856                          */
1857                         if (*ppos + count > inode->i_size)
1858                                 vmtruncate(inode, inode->i_size);
1859                         ret = written;
1860                         goto out_dio;
1861                 }
1862         } else {
1863                 written = generic_file_aio_write_nolock(iocb, iov, nr_segs,
1864                                                         *ppos);
1865         }
1866
1867 out_dio:
1868         /* buffered aio wouldn't have proper lock coverage today */
1869         BUG_ON(ret == -EIOCBQUEUED && !(file->f_flags & O_DIRECT));
1870
1871         if ((file->f_flags & O_SYNC && !direct_io) || IS_SYNC(inode)) {
1872                 /*
1873                  * The generic write paths have handled getting data
1874                  * to disk, but since we don't make use of the dirty
1875                  * inode list, a manual journal commit is necessary
1876                  * here.
1877                  */
1878                 if (old_size != i_size_read(inode) ||
1879                     old_clusters != OCFS2_I(inode)->ip_clusters) {
1880                         ret = jbd2_journal_force_commit(osb->journal->j_journal);
1881                         if (ret < 0)
1882                                 written = ret;
1883                 }
1884         }
1885
1886         /*
1887          * deep in g_f_a_w_n()->ocfs2_direct_IO we pass in a ocfs2_dio_end_io
1888          * function pointer which is called when o_direct io completes so that
1889          * it can unlock our rw lock.  (it's the clustered equivalent of
1890          * i_alloc_sem; protects truncate from racing with pending ios).
1891          * Unfortunately there are error cases which call end_io and others
1892          * that don't.  so we don't have to unlock the rw_lock if either an
1893          * async dio is going to do it in the future or an end_io after an
1894          * error has already done it.
1895          */
1896         if (ret == -EIOCBQUEUED || !ocfs2_iocb_is_rw_locked(iocb)) {
1897                 rw_level = -1;
1898                 have_alloc_sem = 0;
1899         }
1900
1901 out:
1902         if (rw_level != -1)
1903                 ocfs2_rw_unlock(inode, rw_level);
1904
1905 out_sems:
1906         if (have_alloc_sem)
1907                 up_read(&inode->i_alloc_sem);
1908
1909         mutex_unlock(&inode->i_mutex);
1910
1911         mlog_exit(ret);
1912         return written ? written : ret;
1913 }
1914
1915 static int ocfs2_splice_to_file(struct pipe_inode_info *pipe,
1916                                 struct file *out,
1917                                 struct splice_desc *sd)
1918 {
1919         int ret;
1920
1921         ret = ocfs2_prepare_inode_for_write(out->f_path.dentry, &sd->pos,
1922                                             sd->total_len, 0, NULL);
1923         if (ret < 0) {
1924                 mlog_errno(ret);
1925                 return ret;
1926         }
1927
1928         return splice_from_pipe_feed(pipe, sd, pipe_to_file);
1929 }
1930
1931 static ssize_t ocfs2_file_splice_write(struct pipe_inode_info *pipe,
1932                                        struct file *out,
1933                                        loff_t *ppos,
1934                                        size_t len,
1935                                        unsigned int flags)
1936 {
1937         int ret;
1938         struct address_space *mapping = out->f_mapping;
1939         struct inode *inode = mapping->host;
1940         struct splice_desc sd = {
1941                 .total_len = len,
1942                 .flags = flags,
1943                 .pos = *ppos,
1944                 .u.file = out,
1945         };
1946
1947         mlog_entry("(0x%p, 0x%p, %u, '%.*s')\n", out, pipe,
1948                    (unsigned int)len,
1949                    out->f_path.dentry->d_name.len,
1950                    out->f_path.dentry->d_name.name);
1951
1952         if (pipe->inode)
1953                 mutex_lock_nested(&pipe->inode->i_mutex, I_MUTEX_PARENT);
1954
1955         splice_from_pipe_begin(&sd);
1956         do {
1957                 ret = splice_from_pipe_next(pipe, &sd);
1958                 if (ret <= 0)
1959                         break;
1960
1961                 mutex_lock_nested(&inode->i_mutex, I_MUTEX_CHILD);
1962                 ret = ocfs2_rw_lock(inode, 1);
1963                 if (ret < 0)
1964                         mlog_errno(ret);
1965                 else {
1966                         ret = ocfs2_splice_to_file(pipe, out, &sd);
1967                         ocfs2_rw_unlock(inode, 1);
1968                 }
1969                 mutex_unlock(&inode->i_mutex);
1970         } while (ret > 0);
1971         splice_from_pipe_end(pipe, &sd);
1972
1973         if (pipe->inode)
1974                 mutex_unlock(&pipe->inode->i_mutex);
1975
1976         if (sd.num_spliced)
1977                 ret = sd.num_spliced;
1978
1979         if (ret > 0) {
1980                 unsigned long nr_pages;
1981
1982                 *ppos += ret;
1983                 nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1984
1985                 /*
1986                  * If file or inode is SYNC and we actually wrote some data,
1987                  * sync it.
1988                  */
1989                 if (unlikely((out->f_flags & O_SYNC) || IS_SYNC(inode))) {
1990                         int err;
1991
1992                         mutex_lock(&inode->i_mutex);
1993                         err = ocfs2_rw_lock(inode, 1);
1994                         if (err < 0) {
1995                                 mlog_errno(err);
1996                         } else {
1997                                 err = generic_osync_inode(inode, mapping,
1998                                                   OSYNC_METADATA|OSYNC_DATA);
1999                                 ocfs2_rw_unlock(inode, 1);
2000                         }
2001                         mutex_unlock(&inode->i_mutex);
2002
2003                         if (err)
2004                                 ret = err;
2005                 }
2006                 balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
2007         }
2008
2009         mlog_exit(ret);
2010         return ret;
2011 }
2012
2013 static ssize_t ocfs2_file_splice_read(struct file *in,
2014                                       loff_t *ppos,
2015                                       struct pipe_inode_info *pipe,
2016                                       size_t len,
2017                                       unsigned int flags)
2018 {
2019         int ret = 0;
2020         struct inode *inode = in->f_path.dentry->d_inode;
2021
2022         mlog_entry("(0x%p, 0x%p, %u, '%.*s')\n", in, pipe,
2023                    (unsigned int)len,
2024                    in->f_path.dentry->d_name.len,
2025                    in->f_path.dentry->d_name.name);
2026
2027         /*
2028          * See the comment in ocfs2_file_aio_read()
2029          */
2030         ret = ocfs2_inode_lock(inode, NULL, 0);
2031         if (ret < 0) {
2032                 mlog_errno(ret);
2033                 goto bail;
2034         }
2035         ocfs2_inode_unlock(inode, 0);
2036
2037         ret = generic_file_splice_read(in, ppos, pipe, len, flags);
2038
2039 bail:
2040         mlog_exit(ret);
2041         return ret;
2042 }
2043
2044 static ssize_t ocfs2_file_aio_read(struct kiocb *iocb,
2045                                    const struct iovec *iov,
2046                                    unsigned long nr_segs,
2047                                    loff_t pos)
2048 {
2049         int ret = 0, rw_level = -1, have_alloc_sem = 0, lock_level = 0;
2050         struct file *filp = iocb->ki_filp;
2051         struct inode *inode = filp->f_path.dentry->d_inode;
2052
2053         mlog_entry("(0x%p, %u, '%.*s')\n", filp,
2054                    (unsigned int)nr_segs,
2055                    filp->f_path.dentry->d_name.len,
2056                    filp->f_path.dentry->d_name.name);
2057
2058         if (!inode) {
2059                 ret = -EINVAL;
2060                 mlog_errno(ret);
2061                 goto bail;
2062         }
2063
2064         /*
2065          * buffered reads protect themselves in ->readpage().  O_DIRECT reads
2066          * need locks to protect pending reads from racing with truncate.
2067          */
2068         if (filp->f_flags & O_DIRECT) {
2069                 down_read(&inode->i_alloc_sem);
2070                 have_alloc_sem = 1;
2071
2072                 ret = ocfs2_rw_lock(inode, 0);
2073                 if (ret < 0) {
2074                         mlog_errno(ret);
2075                         goto bail;
2076                 }
2077                 rw_level = 0;
2078                 /* communicate with ocfs2_dio_end_io */
2079                 ocfs2_iocb_set_rw_locked(iocb, rw_level);
2080         }
2081
2082         /*
2083          * We're fine letting folks race truncates and extending
2084          * writes with read across the cluster, just like they can
2085          * locally. Hence no rw_lock during read.
2086          *
2087          * Take and drop the meta data lock to update inode fields
2088          * like i_size. This allows the checks down below
2089          * generic_file_aio_read() a chance of actually working.
2090          */
2091         ret = ocfs2_inode_lock_atime(inode, filp->f_vfsmnt, &lock_level);
2092         if (ret < 0) {
2093                 mlog_errno(ret);
2094                 goto bail;
2095         }
2096         ocfs2_inode_unlock(inode, lock_level);
2097
2098         ret = generic_file_aio_read(iocb, iov, nr_segs, iocb->ki_pos);
2099         if (ret == -EINVAL)
2100                 mlog(0, "generic_file_aio_read returned -EINVAL\n");
2101
2102         /* buffered aio wouldn't have proper lock coverage today */
2103         BUG_ON(ret == -EIOCBQUEUED && !(filp->f_flags & O_DIRECT));
2104
2105         /* see ocfs2_file_aio_write */
2106         if (ret == -EIOCBQUEUED || !ocfs2_iocb_is_rw_locked(iocb)) {
2107                 rw_level = -1;
2108                 have_alloc_sem = 0;
2109         }
2110
2111 bail:
2112         if (have_alloc_sem)
2113                 up_read(&inode->i_alloc_sem);
2114         if (rw_level != -1)
2115                 ocfs2_rw_unlock(inode, rw_level);
2116         mlog_exit(ret);
2117
2118         return ret;
2119 }
2120
2121 const struct inode_operations ocfs2_file_iops = {
2122         .setattr        = ocfs2_setattr,
2123         .getattr        = ocfs2_getattr,
2124         .permission     = ocfs2_permission,
2125         .setxattr       = generic_setxattr,
2126         .getxattr       = generic_getxattr,
2127         .listxattr      = ocfs2_listxattr,
2128         .removexattr    = generic_removexattr,
2129         .fallocate      = ocfs2_fallocate,
2130         .fiemap         = ocfs2_fiemap,
2131 };
2132
2133 const struct inode_operations ocfs2_special_file_iops = {
2134         .setattr        = ocfs2_setattr,
2135         .getattr        = ocfs2_getattr,
2136         .permission     = ocfs2_permission,
2137 };
2138
2139 /*
2140  * Other than ->lock, keep ocfs2_fops and ocfs2_dops in sync with
2141  * ocfs2_fops_no_plocks and ocfs2_dops_no_plocks!
2142  */
2143 const struct file_operations ocfs2_fops = {
2144         .llseek         = generic_file_llseek,
2145         .read           = do_sync_read,
2146         .write          = do_sync_write,
2147         .mmap           = ocfs2_mmap,
2148         .fsync          = ocfs2_sync_file,
2149         .release        = ocfs2_file_release,
2150         .open           = ocfs2_file_open,
2151         .aio_read       = ocfs2_file_aio_read,
2152         .aio_write      = ocfs2_file_aio_write,
2153         .unlocked_ioctl = ocfs2_ioctl,
2154 #ifdef CONFIG_COMPAT
2155         .compat_ioctl   = ocfs2_compat_ioctl,
2156 #endif
2157         .lock           = ocfs2_lock,
2158         .flock          = ocfs2_flock,
2159         .splice_read    = ocfs2_file_splice_read,
2160         .splice_write   = ocfs2_file_splice_write,
2161 };
2162
2163 const struct file_operations ocfs2_dops = {
2164         .llseek         = generic_file_llseek,
2165         .read           = generic_read_dir,
2166         .readdir        = ocfs2_readdir,
2167         .fsync          = ocfs2_sync_file,
2168         .release        = ocfs2_dir_release,
2169         .open           = ocfs2_dir_open,
2170         .unlocked_ioctl = ocfs2_ioctl,
2171 #ifdef CONFIG_COMPAT
2172         .compat_ioctl   = ocfs2_compat_ioctl,
2173 #endif
2174         .lock           = ocfs2_lock,
2175         .flock          = ocfs2_flock,
2176 };
2177
2178 /*
2179  * POSIX-lockless variants of our file_operations.
2180  *
2181  * These will be used if the underlying cluster stack does not support
2182  * posix file locking, if the user passes the "localflocks" mount
2183  * option, or if we have a local-only fs.
2184  *
2185  * ocfs2_flock is in here because all stacks handle UNIX file locks,
2186  * so we still want it in the case of no stack support for
2187  * plocks. Internally, it will do the right thing when asked to ignore
2188  * the cluster.
2189  */
2190 const struct file_operations ocfs2_fops_no_plocks = {
2191         .llseek         = generic_file_llseek,
2192         .read           = do_sync_read,
2193         .write          = do_sync_write,
2194         .mmap           = ocfs2_mmap,
2195         .fsync          = ocfs2_sync_file,
2196         .release        = ocfs2_file_release,
2197         .open           = ocfs2_file_open,
2198         .aio_read       = ocfs2_file_aio_read,
2199         .aio_write      = ocfs2_file_aio_write,
2200         .unlocked_ioctl = ocfs2_ioctl,
2201 #ifdef CONFIG_COMPAT
2202         .compat_ioctl   = ocfs2_compat_ioctl,
2203 #endif
2204         .flock          = ocfs2_flock,
2205         .splice_read    = ocfs2_file_splice_read,
2206         .splice_write   = ocfs2_file_splice_write,
2207 };
2208
2209 const struct file_operations ocfs2_dops_no_plocks = {
2210         .llseek         = generic_file_llseek,
2211         .read           = generic_read_dir,
2212         .readdir        = ocfs2_readdir,
2213         .fsync          = ocfs2_sync_file,
2214         .release        = ocfs2_dir_release,
2215         .open           = ocfs2_dir_open,
2216         .unlocked_ioctl = ocfs2_ioctl,
2217 #ifdef CONFIG_COMPAT
2218         .compat_ioctl   = ocfs2_compat_ioctl,
2219 #endif
2220         .flock          = ocfs2_flock,
2221 };