fs/ocfs2/file.c

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