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