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