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/pipe_fs_i.h>
  35 #include <linux/mount.h>
  36 #include <linux/writeback.h>
  37
  38 #define MLOG_MASK_PREFIX ML_INODE
  39 #include <cluster/masklog.h>
  40
  41 #include "ocfs2.h"
  42
  43 #include "alloc.h"
  44 #include "aops.h"
  45 #include "dir.h"
  46 #include "dlmglue.h"
  47 #include "extent_map.h"
  48 #include "file.h"
  49 #include "sysfile.h"
  50 #include "inode.h"
  51 #include "ioctl.h"
  52 #include "journal.h"
  53 #include "mmap.h"
  54 #include "suballoc.h"
  55 #include "super.h"
  56
  57 #include "buffer_head_io.h"
  58
  59 static int ocfs2_sync_inode(struct inode *inode)
  60 {
  61         filemap_fdatawrite(inode->i_mapping);
  62         return sync_mapping_buffers(inode->i_mapping);
  63 }
  64
  65 static int ocfs2_file_open(struct inode *inode, struct file *file)
  66 {
  67         int status;
  68         int mode = file->f_flags;
  69         struct ocfs2_inode_info *oi = OCFS2_I(inode);
  70
  71         mlog_entry("(0x%p, 0x%p, '%.*s')\n", inode, file,
  72                    file->f_path.dentry->d_name.len, file->f_path.dentry->d_name.name);
  73
  74         spin_lock(&oi->ip_lock);
  75
  76         /* Check that the inode hasn't been wiped from disk by another
  77          * node. If it hasn't then we're safe as long as we hold the
  78          * spin lock until our increment of open count. */
  79         if (OCFS2_I(inode)->ip_flags & OCFS2_INODE_DELETED) {
  80                 spin_unlock(&oi->ip_lock);
  81
  82                 status = -ENOENT;
  83                 goto leave;
  84         }
  85
  86         if (mode & O_DIRECT)
  87                 oi->ip_flags |= OCFS2_INODE_OPEN_DIRECT;
  88
  89         oi->ip_open_count++;
  90         spin_unlock(&oi->ip_lock);
  91         status = 0;
  92 leave:
  93         mlog_exit(status);
  94         return status;
  95 }
  96
  97 static int ocfs2_file_release(struct inode *inode, struct file *file)
  98 {
  99         struct ocfs2_inode_info *oi = OCFS2_I(inode);
 100
 101         mlog_entry("(0x%p, 0x%p, '%.*s')\n", inode, file,
 102                        file->f_path.dentry->d_name.len,
 103                        file->f_path.dentry->d_name.name);
 104
 105         spin_lock(&oi->ip_lock);
 106         if (!--oi->ip_open_count)
 107                 oi->ip_flags &= ~OCFS2_INODE_OPEN_DIRECT;
 108         spin_unlock(&oi->ip_lock);
 109
 110         mlog_exit(0);
 111
 112         return 0;
 113 }
 114
 115 static int ocfs2_sync_file(struct file *file,
 116                            struct dentry *dentry,
 117                            int datasync)
 118 {
 119         int err = 0;
 120         journal_t *journal;
 121         struct inode *inode = dentry->d_inode;
 122         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 123
 124         mlog_entry("(0x%p, 0x%p, %d, '%.*s')\n", file, dentry, datasync,
 125                    dentry->d_name.len, dentry->d_name.name);
 126
 127         err = ocfs2_sync_inode(dentry->d_inode);
 128         if (err)
 129                 goto bail;
 130
 131         journal = osb->journal->j_journal;
 132         err = journal_force_commit(journal);
 133
 134 bail:
 135         mlog_exit(err);
 136
 137         return (err < 0) ? -EIO : 0;
 138 }
 139
 140 int ocfs2_should_update_atime(struct inode *inode,
 141                               struct vfsmount *vfsmnt)
 142 {
 143         struct timespec now;
 144         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 145
 146         if (ocfs2_is_hard_readonly(osb) || ocfs2_is_soft_readonly(osb))
 147                 return 0;
 148
 149         if ((inode->i_flags & S_NOATIME) ||
 150             ((inode->i_sb->s_flags & MS_NODIRATIME) && S_ISDIR(inode->i_mode)))
 151                 return 0;
 152
 153         /*
 154          * We can be called with no vfsmnt structure - NFSD will
 155          * sometimes do this.
 156          *
 157          * Note that our action here is different than touch_atime() -
 158          * if we can't tell whether this is a noatime mount, then we
 159          * don't know whether to trust the value of s_atime_quantum.
 160          */
 161         if (vfsmnt == NULL)
 162                 return 0;
 163
 164         if ((vfsmnt->mnt_flags & MNT_NOATIME) ||
 165             ((vfsmnt->mnt_flags & MNT_NODIRATIME) && S_ISDIR(inode->i_mode)))
 166                 return 0;
 167
 168         if (vfsmnt->mnt_flags & MNT_RELATIME) {
 169                 if ((timespec_compare(&inode->i_atime, &inode->i_mtime) <= 0) ||
 170                     (timespec_compare(&inode->i_atime, &inode->i_ctime) <= 0))
 171                         return 1;
 172
 173                 return 0;
 174         }
 175
 176         now = CURRENT_TIME;
 177         if ((now.tv_sec - inode->i_atime.tv_sec <= osb->s_atime_quantum))
 178                 return 0;
 179         else
 180                 return 1;
 181 }
 182
 183 int ocfs2_update_inode_atime(struct inode *inode,
 184                              struct buffer_head *bh)
 185 {
 186         int ret;
 187         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 188         handle_t *handle;
 189
 190         mlog_entry_void();
 191
 192         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
 193         if (handle == NULL) {
 194                 ret = -ENOMEM;
 195                 mlog_errno(ret);
 196                 goto out;
 197         }
 198
 199         inode->i_atime = CURRENT_TIME;
 200         ret = ocfs2_mark_inode_dirty(handle, inode, bh);
 201         if (ret < 0)
 202                 mlog_errno(ret);
 203
 204         ocfs2_commit_trans(OCFS2_SB(inode->i_sb), handle);
 205 out:
 206         mlog_exit(ret);
 207         return ret;
 208 }
 209
 210 static int ocfs2_set_inode_size(handle_t *handle,
 211                                 struct inode *inode,
 212                                 struct buffer_head *fe_bh,
 213                                 u64 new_i_size)
 214 {
 215         int status;
 216
 217         mlog_entry_void();
 218         i_size_write(inode, new_i_size);
 219         inode->i_blocks = ocfs2_inode_sector_count(inode);
 220         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
 221
 222         status = ocfs2_mark_inode_dirty(handle, inode, fe_bh);
 223         if (status < 0) {
 224                 mlog_errno(status);
 225                 goto bail;
 226         }
 227
 228 bail:
 229         mlog_exit(status);
 230         return status;
 231 }
 232
 233 static int ocfs2_simple_size_update(struct inode *inode,
 234                                     struct buffer_head *di_bh,
 235                                     u64 new_i_size)
 236 {
 237         int ret;
 238         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 239         handle_t *handle = NULL;
 240
 241         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
 242         if (handle == NULL) {
 243                 ret = -ENOMEM;
 244                 mlog_errno(ret);
 245                 goto out;
 246         }
 247
 248         ret = ocfs2_set_inode_size(handle, inode, di_bh,
 249                                    new_i_size);
 250         if (ret < 0)
 251                 mlog_errno(ret);
 252
 253         ocfs2_commit_trans(osb, handle);
 254 out:
 255         return ret;
 256 }
 257
 258 static int ocfs2_orphan_for_truncate(struct ocfs2_super *osb,
 259                                      struct inode *inode,
 260                                      struct buffer_head *fe_bh,
 261                                      u64 new_i_size)
 262 {
 263         int status;
 264         handle_t *handle;
 265         struct ocfs2_dinode *di;
 266
 267         mlog_entry_void();
 268
 269         /* TODO: This needs to actually orphan the inode in this
 270          * transaction. */
 271
 272         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
 273         if (IS_ERR(handle)) {
 274                 status = PTR_ERR(handle);
 275                 mlog_errno(status);
 276                 goto out;
 277         }
 278
 279         status = ocfs2_journal_access(handle, inode, fe_bh,
 280                                       OCFS2_JOURNAL_ACCESS_WRITE);
 281         if (status < 0) {
 282                 mlog_errno(status);
 283                 goto out_commit;
 284         }
 285
 286         /*
 287          * Do this before setting i_size.
 288          */
 289         status = ocfs2_zero_tail_for_truncate(inode, handle, new_i_size);
 290         if (status) {
 291                 mlog_errno(status);
 292                 goto out_commit;
 293         }
 294
 295         i_size_write(inode, new_i_size);
 296         inode->i_blocks = ocfs2_align_bytes_to_sectors(new_i_size);
 297         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
 298
 299         di = (struct ocfs2_dinode *) fe_bh->b_data;
 300         di->i_size = cpu_to_le64(new_i_size);
 301         di->i_ctime = di->i_mtime = cpu_to_le64(inode->i_ctime.tv_sec);
 302         di->i_ctime_nsec = di->i_mtime_nsec = cpu_to_le32(inode->i_ctime.tv_nsec);
 303
 304         status = ocfs2_journal_dirty(handle, fe_bh);
 305         if (status < 0)
 306                 mlog_errno(status);
 307
 308 out_commit:
 309         ocfs2_commit_trans(osb, handle);
 310 out:
 311
 312         mlog_exit(status);
 313         return status;
 314 }
 315
 316 static int ocfs2_truncate_file(struct inode *inode,
 317                                struct buffer_head *di_bh,
 318                                u64 new_i_size)
 319 {
 320         int status = 0;
 321         struct ocfs2_dinode *fe = NULL;
 322         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 323         struct ocfs2_truncate_context *tc = NULL;
 324
 325         mlog_entry("(inode = %llu, new_i_size = %llu\n",
 326                    (unsigned long long)OCFS2_I(inode)->ip_blkno,
 327                    (unsigned long long)new_i_size);
 328
 329         truncate_inode_pages(inode->i_mapping, new_i_size);
 330
 331         fe = (struct ocfs2_dinode *) di_bh->b_data;
 332         if (!OCFS2_IS_VALID_DINODE(fe)) {
 333                 OCFS2_RO_ON_INVALID_DINODE(inode->i_sb, fe);
 334                 status = -EIO;
 335                 goto bail;
 336         }
 337
 338         mlog_bug_on_msg(le64_to_cpu(fe->i_size) != i_size_read(inode),
 339                         "Inode %llu, inode i_size = %lld != di "
 340                         "i_size = %llu, i_flags = 0x%x\n",
 341                         (unsigned long long)OCFS2_I(inode)->ip_blkno,
 342                         i_size_read(inode),
 343                         (unsigned long long)le64_to_cpu(fe->i_size),
 344                         le32_to_cpu(fe->i_flags));
 345
 346         if (new_i_size > le64_to_cpu(fe->i_size)) {
 347                 mlog(0, "asked to truncate file with size (%llu) to size (%llu)!\n",
 348                      (unsigned long long)le64_to_cpu(fe->i_size),
 349                      (unsigned long long)new_i_size);
 350                 status = -EINVAL;
 351                 mlog_errno(status);
 352                 goto bail;
 353         }
 354
 355         mlog(0, "inode %llu, i_size = %llu, new_i_size = %llu\n",
 356              (unsigned long long)le64_to_cpu(fe->i_blkno),
 357              (unsigned long long)le64_to_cpu(fe->i_size),
 358              (unsigned long long)new_i_size);
 359
 360         /* lets handle the simple truncate cases before doing any more
 361          * cluster locking. */
 362         if (new_i_size == le64_to_cpu(fe->i_size))
 363                 goto bail;
 364
 365         /* This forces other nodes to sync and drop their pages. Do
 366          * this even if we have a truncate without allocation change -
 367          * ocfs2 cluster sizes can be much greater than page size, so
 368          * we have to truncate them anyway.  */
 369         status = ocfs2_data_lock(inode, 1);
 370         if (status < 0) {
 371                 mlog_errno(status);
 372                 goto bail;
 373         }
 374
 375         /* alright, we're going to need to do a full blown alloc size
 376          * change. Orphan the inode so that recovery can complete the
 377          * truncate if necessary. This does the task of marking
 378          * i_size. */
 379         status = ocfs2_orphan_for_truncate(osb, inode, di_bh, new_i_size);
 380         if (status < 0) {
 381                 mlog_errno(status);
 382                 goto bail_unlock_data;
 383         }
 384
 385         status = ocfs2_prepare_truncate(osb, inode, di_bh, &tc);
 386         if (status < 0) {
 387                 mlog_errno(status);
 388                 goto bail_unlock_data;
 389         }
 390
 391         status = ocfs2_commit_truncate(osb, inode, di_bh, tc);
 392         if (status < 0) {
 393                 mlog_errno(status);
 394                 goto bail_unlock_data;
 395         }
 396
 397         /* TODO: orphan dir cleanup here. */
 398 bail_unlock_data:
 399         ocfs2_data_unlock(inode, 1);
 400
 401 bail:
 402
 403         mlog_exit(status);
 404         return status;
 405 }
 406
 407 /*
 408  * extend allocation only here.
 409  * we'll update all the disk stuff, and oip->alloc_size
 410  *
 411  * expect stuff to be locked, a transaction started and enough data /
 412  * metadata reservations in the contexts.
 413  *
 414  * Will return -EAGAIN, and a reason if a restart is needed.
 415  * If passed in, *reason will always be set, even in error.
 416  */
 417 int ocfs2_do_extend_allocation(struct ocfs2_super *osb,
 418                                struct inode *inode,
 419                                u32 *logical_offset,
 420                                u32 clusters_to_add,
 421                                struct buffer_head *fe_bh,
 422                                handle_t *handle,
 423                                struct ocfs2_alloc_context *data_ac,
 424                                struct ocfs2_alloc_context *meta_ac,
 425                                enum ocfs2_alloc_restarted *reason_ret)
 426 {
 427         int status = 0;
 428         int free_extents;
 429         struct ocfs2_dinode *fe = (struct ocfs2_dinode *) fe_bh->b_data;
 430         enum ocfs2_alloc_restarted reason = RESTART_NONE;
 431         u32 bit_off, num_bits;
 432         u64 block;
 433
 434         BUG_ON(!clusters_to_add);
 435
 436         free_extents = ocfs2_num_free_extents(osb, inode, fe);
 437         if (free_extents < 0) {
 438                 status = free_extents;
 439                 mlog_errno(status);
 440                 goto leave;
 441         }
 442
 443         /* there are two cases which could cause us to EAGAIN in the
 444          * we-need-more-metadata case:
 445          * 1) we haven't reserved *any*
 446          * 2) we are so fragmented, we've needed to add metadata too
 447          *    many times. */
 448         if (!free_extents && !meta_ac) {
 449                 mlog(0, "we haven't reserved any metadata!\n");
 450                 status = -EAGAIN;
 451                 reason = RESTART_META;
 452                 goto leave;
 453         } else if ((!free_extents)
 454                    && (ocfs2_alloc_context_bits_left(meta_ac)
 455                        < ocfs2_extend_meta_needed(fe))) {
 456                 mlog(0, "filesystem is really fragmented...\n");
 457                 status = -EAGAIN;
 458                 reason = RESTART_META;
 459                 goto leave;
 460         }
 461
 462         status = ocfs2_claim_clusters(osb, handle, data_ac, 1,
 463                                       &bit_off, &num_bits);
 464         if (status < 0) {
 465                 if (status != -ENOSPC)
 466                         mlog_errno(status);
 467                 goto leave;
 468         }
 469
 470         BUG_ON(num_bits > clusters_to_add);
 471
 472         /* reserve our write early -- insert_extent may update the inode */
 473         status = ocfs2_journal_access(handle, inode, fe_bh,
 474                                       OCFS2_JOURNAL_ACCESS_WRITE);
 475         if (status < 0) {
 476                 mlog_errno(status);
 477                 goto leave;
 478         }
 479
 480         block = ocfs2_clusters_to_blocks(osb->sb, bit_off);
 481         mlog(0, "Allocating %u clusters at block %u for inode %llu\n",
 482              num_bits, bit_off, (unsigned long long)OCFS2_I(inode)->ip_blkno);
 483         status = ocfs2_insert_extent(osb, handle, inode, fe_bh,
 484                                      *logical_offset, block, num_bits,
 485                                      meta_ac);
 486         if (status < 0) {
 487                 mlog_errno(status);
 488                 goto leave;
 489         }
 490
 491         status = ocfs2_journal_dirty(handle, fe_bh);
 492         if (status < 0) {
 493                 mlog_errno(status);
 494                 goto leave;
 495         }
 496
 497         clusters_to_add -= num_bits;
 498         *logical_offset += num_bits;
 499
 500         if (clusters_to_add) {
 501                 mlog(0, "need to alloc once more, clusters = %u, wanted = "
 502                      "%u\n", fe->i_clusters, clusters_to_add);
 503                 status = -EAGAIN;
 504                 reason = RESTART_TRANS;
 505         }
 506
 507 leave:
 508         mlog_exit(status);
 509         if (reason_ret)
 510                 *reason_ret = reason;
 511         return status;
 512 }
 513
 514 /*
 515  * For a given allocation, determine which allocators will need to be
 516  * accessed, and lock them, reserving the appropriate number of bits.
 517  *
 518  * Called from ocfs2_extend_allocation() for file systems which don't
 519  * support holes, and from ocfs2_write() for file systems which
 520  * understand sparse inodes.
 521  */
 522 int ocfs2_lock_allocators(struct inode *inode, struct ocfs2_dinode *di,
 523                           u32 clusters_to_add,
 524                           struct ocfs2_alloc_context **data_ac,
 525                           struct ocfs2_alloc_context **meta_ac)
 526 {
 527         int ret, num_free_extents;
 528         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 529
 530         *meta_ac = NULL;
 531         *data_ac = NULL;
 532
 533         mlog(0, "extend inode %llu, i_size = %lld, di->i_clusters = %u, "
 534              "clusters_to_add = %u\n",
 535              (unsigned long long)OCFS2_I(inode)->ip_blkno, i_size_read(inode),
 536              le32_to_cpu(di->i_clusters), clusters_to_add);
 537
 538         num_free_extents = ocfs2_num_free_extents(osb, inode, di);
 539         if (num_free_extents < 0) {
 540                 ret = num_free_extents;
 541                 mlog_errno(ret);
 542                 goto out;
 543         }
 544
 545         /*
 546          * Sparse allocation file systems need to be more conservative
 547          * with reserving room for expansion - the actual allocation
 548          * happens while we've got a journal handle open so re-taking
 549          * a cluster lock (because we ran out of room for another
 550          * extent) will violate ordering rules.
 551          *
 552          * Most of the time we'll only be seeing this 1 cluster at a time
 553          * anyway.
 554          */
 555         if (!num_free_extents ||
 556             (ocfs2_sparse_alloc(osb) && num_free_extents < clusters_to_add)) {
 557                 ret = ocfs2_reserve_new_metadata(osb, di, meta_ac);
 558                 if (ret < 0) {
 559                         if (ret != -ENOSPC)
 560                                 mlog_errno(ret);
 561                         goto out;
 562                 }
 563         }
 564
 565         ret = ocfs2_reserve_clusters(osb, clusters_to_add, data_ac);
 566         if (ret < 0) {
 567                 if (ret != -ENOSPC)
 568                         mlog_errno(ret);
 569                 goto out;
 570         }
 571
 572 out:
 573         if (ret) {
 574                 if (*meta_ac) {
 575                         ocfs2_free_alloc_context(*meta_ac);
 576                         *meta_ac = NULL;
 577                 }
 578
 579                 /*
 580                  * We cannot have an error and a non null *data_ac.
 581                  */
 582         }
 583
 584         return ret;
 585 }
 586
 587 static int ocfs2_extend_allocation(struct inode *inode,
 588                                    u32 clusters_to_add)
 589 {
 590         int status = 0;
 591         int restart_func = 0;
 592         int drop_alloc_sem = 0;
 593         int credits;
 594         u32 prev_clusters, logical_start;
 595         struct buffer_head *bh = NULL;
 596         struct ocfs2_dinode *fe = NULL;
 597         handle_t *handle = NULL;
 598         struct ocfs2_alloc_context *data_ac = NULL;
 599         struct ocfs2_alloc_context *meta_ac = NULL;
 600         enum ocfs2_alloc_restarted why;
 601         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 602
 603         mlog_entry("(clusters_to_add = %u)\n", clusters_to_add);
 604
 605         /*
 606          * This function only exists for file systems which don't
 607          * support holes.
 608          */
 609         BUG_ON(ocfs2_sparse_alloc(osb));
 610
 611         status = ocfs2_read_block(osb, OCFS2_I(inode)->ip_blkno, &bh,
 612                                   OCFS2_BH_CACHED, inode);
 613         if (status < 0) {
 614                 mlog_errno(status);
 615                 goto leave;
 616         }
 617
 618         fe = (struct ocfs2_dinode *) bh->b_data;
 619         if (!OCFS2_IS_VALID_DINODE(fe)) {
 620                 OCFS2_RO_ON_INVALID_DINODE(inode->i_sb, fe);
 621                 status = -EIO;
 622                 goto leave;
 623         }
 624
 625         logical_start = OCFS2_I(inode)->ip_clusters;
 626
 627 restart_all:
 628         BUG_ON(le32_to_cpu(fe->i_clusters) != OCFS2_I(inode)->ip_clusters);
 629
 630         /* blocks peope in read/write from reading our allocation
 631          * until we're done changing it. We depend on i_mutex to block
 632          * other extend/truncate calls while we're here. Ordering wrt
 633          * start_trans is important here -- always do it before! */
 634         down_write(&OCFS2_I(inode)->ip_alloc_sem);
 635         drop_alloc_sem = 1;
 636
 637         status = ocfs2_lock_allocators(inode, fe, clusters_to_add, &data_ac,
 638                                        &meta_ac);
 639         if (status) {
 640                 mlog_errno(status);
 641                 goto leave;
 642         }
 643
 644         credits = ocfs2_calc_extend_credits(osb->sb, fe, clusters_to_add);
 645         handle = ocfs2_start_trans(osb, credits);
 646         if (IS_ERR(handle)) {
 647                 status = PTR_ERR(handle);
 648                 handle = NULL;
 649                 mlog_errno(status);
 650                 goto leave;
 651         }
 652
 653 restarted_transaction:
 654         /* reserve a write to the file entry early on - that we if we
 655          * run out of credits in the allocation path, we can still
 656          * update i_size. */
 657         status = ocfs2_journal_access(handle, inode, bh,
 658                                       OCFS2_JOURNAL_ACCESS_WRITE);
 659         if (status < 0) {
 660                 mlog_errno(status);
 661                 goto leave;
 662         }
 663
 664         prev_clusters = OCFS2_I(inode)->ip_clusters;
 665
 666         status = ocfs2_do_extend_allocation(osb,
 667                                             inode,
 668                                             &logical_start,
 669                                             clusters_to_add,
 670                                             bh,
 671                                             handle,
 672                                             data_ac,
 673                                             meta_ac,
 674                                             &why);
 675         if ((status < 0) && (status != -EAGAIN)) {
 676                 if (status != -ENOSPC)
 677                         mlog_errno(status);
 678                 goto leave;
 679         }
 680
 681         status = ocfs2_journal_dirty(handle, bh);
 682         if (status < 0) {
 683                 mlog_errno(status);
 684                 goto leave;
 685         }
 686
 687         spin_lock(&OCFS2_I(inode)->ip_lock);
 688         clusters_to_add -= (OCFS2_I(inode)->ip_clusters - prev_clusters);
 689         spin_unlock(&OCFS2_I(inode)->ip_lock);
 690
 691         if (why != RESTART_NONE && clusters_to_add) {
 692                 if (why == RESTART_META) {
 693                         mlog(0, "restarting function.\n");
 694                         restart_func = 1;
 695                 } else {
 696                         BUG_ON(why != RESTART_TRANS);
 697
 698                         mlog(0, "restarting transaction.\n");
 699                         /* TODO: This can be more intelligent. */
 700                         credits = ocfs2_calc_extend_credits(osb->sb,
 701                                                             fe,
 702                                                             clusters_to_add);
 703                         status = ocfs2_extend_trans(handle, credits);
 704                         if (status < 0) {
 705                                 /* handle still has to be committed at
 706                                  * this point. */
 707                                 status = -ENOMEM;
 708                                 mlog_errno(status);
 709                                 goto leave;
 710                         }
 711                         goto restarted_transaction;
 712                 }
 713         }
 714
 715         mlog(0, "fe: i_clusters = %u, i_size=%llu\n",
 716              le32_to_cpu(fe->i_clusters),
 717              (unsigned long long)le64_to_cpu(fe->i_size));
 718         mlog(0, "inode: ip_clusters=%u, i_size=%lld\n",
 719              OCFS2_I(inode)->ip_clusters, i_size_read(inode));
 720
 721 leave:
 722         if (drop_alloc_sem) {
 723                 up_write(&OCFS2_I(inode)->ip_alloc_sem);
 724                 drop_alloc_sem = 0;
 725         }
 726         if (handle) {
 727                 ocfs2_commit_trans(osb, handle);
 728                 handle = NULL;
 729         }
 730         if (data_ac) {
 731                 ocfs2_free_alloc_context(data_ac);
 732                 data_ac = NULL;
 733         }
 734         if (meta_ac) {
 735                 ocfs2_free_alloc_context(meta_ac);
 736                 meta_ac = NULL;
 737         }
 738         if ((!status) && restart_func) {
 739                 restart_func = 0;
 740                 goto restart_all;
 741         }
 742         if (bh) {
 743                 brelse(bh);
 744                 bh = NULL;
 745         }
 746
 747         mlog_exit(status);
 748         return status;
 749 }
 750
 751 /* Some parts of this taken from generic_cont_expand, which turned out
 752  * to be too fragile to do exactly what we need without us having to
 753  * worry about recursive locking in ->prepare_write() and
 754  * ->commit_write(). */
 755 static int ocfs2_write_zero_page(struct inode *inode,
 756                                  u64 size)
 757 {
 758         struct address_space *mapping = inode->i_mapping;
 759         struct page *page;
 760         unsigned long index;
 761         unsigned int offset;
 762         handle_t *handle = NULL;
 763         int ret;
 764
 765         offset = (size & (PAGE_CACHE_SIZE-1)); /* Within page */
 766         /* ugh.  in prepare/commit_write, if from==to==start of block, we
 767         ** skip the prepare.  make sure we never send an offset for the start
 768         ** of a block
 769         */
 770         if ((offset & (inode->i_sb->s_blocksize - 1)) == 0) {
 771                 offset++;
 772         }
 773         index = size >> PAGE_CACHE_SHIFT;
 774
 775         page = grab_cache_page(mapping, index);
 776         if (!page) {
 777                 ret = -ENOMEM;
 778                 mlog_errno(ret);
 779                 goto out;
 780         }
 781
 782         ret = ocfs2_prepare_write_nolock(inode, page, offset, offset);
 783         if (ret < 0) {
 784                 mlog_errno(ret);
 785                 goto out_unlock;
 786         }
 787
 788         if (ocfs2_should_order_data(inode)) {
 789                 handle = ocfs2_start_walk_page_trans(inode, page, offset,
 790                                                      offset);
 791                 if (IS_ERR(handle)) {
 792                         ret = PTR_ERR(handle);
 793                         handle = NULL;
 794                         goto out_unlock;
 795                 }
 796         }
 797
 798         /* must not update i_size! */
 799         ret = block_commit_write(page, offset, offset);
 800         if (ret < 0)
 801                 mlog_errno(ret);
 802         else
 803                 ret = 0;
 804
 805         if (handle)
 806                 ocfs2_commit_trans(OCFS2_SB(inode->i_sb), handle);
 807 out_unlock:
 808         unlock_page(page);
 809         page_cache_release(page);
 810 out:
 811         return ret;
 812 }
 813
 814 static int ocfs2_zero_extend(struct inode *inode,
 815                              u64 zero_to_size)
 816 {
 817         int ret = 0;
 818         u64 start_off;
 819         struct super_block *sb = inode->i_sb;
 820
 821         start_off = ocfs2_align_bytes_to_blocks(sb, i_size_read(inode));
 822         while (start_off < zero_to_size) {
 823                 ret = ocfs2_write_zero_page(inode, start_off);
 824                 if (ret < 0) {
 825                         mlog_errno(ret);
 826                         goto out;
 827                 }
 828
 829                 start_off += sb->s_blocksize;
 830
 831                 /*
 832                  * Very large extends have the potential to lock up
 833                  * the cpu for extended periods of time.
 834                  */
 835                 cond_resched();
 836         }
 837
 838 out:
 839         return ret;
 840 }
 841
 842 /*
 843  * A tail_to_skip value > 0 indicates that we're being called from
 844  * ocfs2_file_aio_write(). This has the following implications:
 845  *
 846  * - we don't want to update i_size
 847  * - di_bh will be NULL, which is fine because it's only used in the
 848  *   case where we want to update i_size.
 849  * - ocfs2_zero_extend() will then only be filling the hole created
 850  *   between i_size and the start of the write.
 851  */
 852 static int ocfs2_extend_file(struct inode *inode,
 853                              struct buffer_head *di_bh,
 854                              u64 new_i_size,
 855                              size_t tail_to_skip)
 856 {
 857         int ret = 0;
 858         u32 clusters_to_add = 0;
 859
 860         BUG_ON(!tail_to_skip && !di_bh);
 861
 862         /* setattr sometimes calls us like this. */
 863         if (new_i_size == 0)
 864                 goto out;
 865
 866         if (i_size_read(inode) == new_i_size)
 867                 goto out;
 868         BUG_ON(new_i_size < i_size_read(inode));
 869
 870         if (ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb))) {
 871                 BUG_ON(tail_to_skip != 0);
 872                 goto out_update_size;
 873         }
 874
 875         clusters_to_add = ocfs2_clusters_for_bytes(inode->i_sb, new_i_size) -
 876                 OCFS2_I(inode)->ip_clusters;
 877
 878         /*
 879          * protect the pages that ocfs2_zero_extend is going to be
 880          * pulling into the page cache.. we do this before the
 881          * metadata extend so that we don't get into the situation
 882          * where we've extended the metadata but can't get the data
 883          * lock to zero.
 884          */
 885         ret = ocfs2_data_lock(inode, 1);
 886         if (ret < 0) {
 887                 mlog_errno(ret);
 888                 goto out;
 889         }
 890
 891         if (clusters_to_add) {
 892                 ret = ocfs2_extend_allocation(inode, clusters_to_add);
 893                 if (ret < 0) {
 894                         mlog_errno(ret);
 895                         goto out_unlock;
 896                 }
 897         }
 898
 899         /*
 900          * Call this even if we don't add any clusters to the tree. We
 901          * still need to zero the area between the old i_size and the
 902          * new i_size.
 903          */
 904         ret = ocfs2_zero_extend(inode, (u64)new_i_size - tail_to_skip);
 905         if (ret < 0) {
 906                 mlog_errno(ret);
 907                 goto out_unlock;
 908         }
 909
 910 out_update_size:
 911         if (!tail_to_skip) {
 912                 /* We're being called from ocfs2_setattr() which wants
 913                  * us to update i_size */
 914                 ret = ocfs2_simple_size_update(inode, di_bh, new_i_size);
 915                 if (ret < 0)
 916                         mlog_errno(ret);
 917         }
 918
 919 out_unlock:
 920         if (!ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)))
 921                 ocfs2_data_unlock(inode, 1);
 922
 923 out:
 924         return ret;
 925 }
 926
 927 int ocfs2_setattr(struct dentry *dentry, struct iattr *attr)
 928 {
 929         int status = 0, size_change;
 930         struct inode *inode = dentry->d_inode;
 931         struct super_block *sb = inode->i_sb;
 932         struct ocfs2_super *osb = OCFS2_SB(sb);
 933         struct buffer_head *bh = NULL;
 934         handle_t *handle = NULL;
 935
 936         mlog_entry("(0x%p, '%.*s')\n", dentry,
 937                    dentry->d_name.len, dentry->d_name.name);
 938
 939         if (attr->ia_valid & ATTR_MODE)
 940                 mlog(0, "mode change: %d\n", attr->ia_mode);
 941         if (attr->ia_valid & ATTR_UID)
 942                 mlog(0, "uid change: %d\n", attr->ia_uid);
 943         if (attr->ia_valid & ATTR_GID)
 944                 mlog(0, "gid change: %d\n", attr->ia_gid);
 945         if (attr->ia_valid & ATTR_SIZE)
 946                 mlog(0, "size change...\n");
 947         if (attr->ia_valid & (ATTR_ATIME | ATTR_MTIME | ATTR_CTIME))
 948                 mlog(0, "time change...\n");
 949
 950 #define OCFS2_VALID_ATTRS (ATTR_ATIME | ATTR_MTIME | ATTR_CTIME | ATTR_SIZE \
 951                            | ATTR_GID | ATTR_UID | ATTR_MODE)
 952         if (!(attr->ia_valid & OCFS2_VALID_ATTRS)) {
 953                 mlog(0, "can't handle attrs: 0x%x\n", attr->ia_valid);
 954                 return 0;
 955         }
 956
 957         status = inode_change_ok(inode, attr);
 958         if (status)
 959                 return status;
 960
 961         size_change = S_ISREG(inode->i_mode) && attr->ia_valid & ATTR_SIZE;
 962         if (size_change) {
 963                 status = ocfs2_rw_lock(inode, 1);
 964                 if (status < 0) {
 965                         mlog_errno(status);
 966                         goto bail;
 967                 }
 968         }
 969
 970         status = ocfs2_meta_lock(inode, &bh, 1);
 971         if (status < 0) {
 972                 if (status != -ENOENT)
 973                         mlog_errno(status);
 974                 goto bail_unlock_rw;
 975         }
 976
 977         if (size_change && attr->ia_size != i_size_read(inode)) {
 978                 if (i_size_read(inode) > attr->ia_size)
 979                         status = ocfs2_truncate_file(inode, bh, attr->ia_size);
 980                 else
 981                         status = ocfs2_extend_file(inode, bh, attr->ia_size, 0);
 982                 if (status < 0) {
 983                         if (status != -ENOSPC)
 984                                 mlog_errno(status);
 985                         status = -ENOSPC;
 986                         goto bail_unlock;
 987                 }
 988         }
 989
 990         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
 991         if (IS_ERR(handle)) {
 992                 status = PTR_ERR(handle);
 993                 mlog_errno(status);
 994                 goto bail_unlock;
 995         }
 996
 997         status = inode_setattr(inode, attr);
 998         if (status < 0) {
 999                 mlog_errno(status);
1000                 goto bail_commit;
1001         }
1002
1003         status = ocfs2_mark_inode_dirty(handle, inode, bh);
1004         if (status < 0)
1005                 mlog_errno(status);
1006
1007 bail_commit:
1008         ocfs2_commit_trans(osb, handle);
1009 bail_unlock:
1010         ocfs2_meta_unlock(inode, 1);
1011 bail_unlock_rw:
1012         if (size_change)
1013                 ocfs2_rw_unlock(inode, 1);
1014 bail:
1015         if (bh)
1016                 brelse(bh);
1017
1018         mlog_exit(status);
1019         return status;
1020 }
1021
1022 int ocfs2_getattr(struct vfsmount *mnt,
1023                   struct dentry *dentry,
1024                   struct kstat *stat)
1025 {
1026         struct inode *inode = dentry->d_inode;
1027         struct super_block *sb = dentry->d_inode->i_sb;
1028         struct ocfs2_super *osb = sb->s_fs_info;
1029         int err;
1030
1031         mlog_entry_void();
1032
1033         err = ocfs2_inode_revalidate(dentry);
1034         if (err) {
1035                 if (err != -ENOENT)
1036                         mlog_errno(err);
1037                 goto bail;
1038         }
1039
1040         generic_fillattr(inode, stat);
1041
1042         /* We set the blksize from the cluster size for performance */
1043         stat->blksize = osb->s_clustersize;
1044
1045 bail:
1046         mlog_exit(err);
1047
1048         return err;
1049 }
1050
1051 int ocfs2_permission(struct inode *inode, int mask, struct nameidata *nd)
1052 {
1053         int ret;
1054
1055         mlog_entry_void();
1056
1057         ret = ocfs2_meta_lock(inode, NULL, 0);
1058         if (ret) {
1059                 if (ret != -ENOENT)
1060                         mlog_errno(ret);
1061                 goto out;
1062         }
1063
1064         ret = generic_permission(inode, mask, NULL);
1065
1066         ocfs2_meta_unlock(inode, 0);
1067 out:
1068         mlog_exit(ret);
1069         return ret;
1070 }
1071
1072 static int ocfs2_write_remove_suid(struct inode *inode)
1073 {
1074         int ret;
1075         struct buffer_head *bh = NULL;
1076         struct ocfs2_inode_info *oi = OCFS2_I(inode);
1077         handle_t *handle;
1078         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1079         struct ocfs2_dinode *di;
1080
1081         mlog_entry("(Inode %llu, mode 0%o)\n",
1082                    (unsigned long long)oi->ip_blkno, inode->i_mode);
1083
1084         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1085         if (handle == NULL) {
1086                 ret = -ENOMEM;
1087                 mlog_errno(ret);
1088                 goto out;
1089         }
1090
1091         ret = ocfs2_read_block(osb, oi->ip_blkno, &bh, OCFS2_BH_CACHED, inode);
1092         if (ret < 0) {
1093                 mlog_errno(ret);
1094                 goto out_trans;
1095         }
1096
1097         ret = ocfs2_journal_access(handle, inode, bh,
1098                                    OCFS2_JOURNAL_ACCESS_WRITE);
1099         if (ret < 0) {
1100                 mlog_errno(ret);
1101                 goto out_bh;
1102         }
1103
1104         inode->i_mode &= ~S_ISUID;
1105         if ((inode->i_mode & S_ISGID) && (inode->i_mode & S_IXGRP))
1106                 inode->i_mode &= ~S_ISGID;
1107
1108         di = (struct ocfs2_dinode *) bh->b_data;
1109         di->i_mode = cpu_to_le16(inode->i_mode);
1110
1111         ret = ocfs2_journal_dirty(handle, bh);
1112         if (ret < 0)
1113                 mlog_errno(ret);
1114 out_bh:
1115         brelse(bh);
1116 out_trans:
1117         ocfs2_commit_trans(osb, handle);
1118 out:
1119         mlog_exit(ret);
1120         return ret;
1121 }
1122
1123 /*
1124  * Will look for holes and unwritten extents in the range starting at
1125  * pos for count bytes (inclusive).
1126  */
1127 static int ocfs2_check_range_for_holes(struct inode *inode, loff_t pos,
1128                                        size_t count)
1129 {
1130         int ret = 0;
1131         unsigned int extent_flags;
1132         u32 cpos, clusters, extent_len, phys_cpos;
1133         struct super_block *sb = inode->i_sb;
1134
1135         cpos = pos >> OCFS2_SB(sb)->s_clustersize_bits;
1136         clusters = ocfs2_clusters_for_bytes(sb, pos + count) - cpos;
1137
1138         while (clusters) {
1139                 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos, &extent_len,
1140                                          &extent_flags);
1141                 if (ret < 0) {
1142                         mlog_errno(ret);
1143                         goto out;
1144                 }
1145
1146                 if (phys_cpos == 0 || (extent_flags & OCFS2_EXT_UNWRITTEN)) {
1147                         ret = 1;
1148                         break;
1149                 }
1150
1151                 if (extent_len > clusters)
1152                         extent_len = clusters;
1153
1154                 clusters -= extent_len;
1155                 cpos += extent_len;
1156         }
1157 out:
1158         return ret;
1159 }
1160
1161 static int ocfs2_prepare_inode_for_write(struct dentry *dentry,
1162                                          loff_t *ppos,
1163                                          size_t count,
1164                                          int appending,
1165                                          int *direct_io)
1166 {
1167         int ret = 0, meta_level = appending;
1168         struct inode *inode = dentry->d_inode;
1169         u32 clusters;
1170         loff_t newsize, saved_pos;
1171
1172         /*
1173          * We sample i_size under a read level meta lock to see if our write
1174          * is extending the file, if it is we back off and get a write level
1175          * meta lock.
1176          */
1177         for(;;) {
1178                 ret = ocfs2_meta_lock(inode, NULL, meta_level);
1179                 if (ret < 0) {
1180                         meta_level = -1;
1181                         mlog_errno(ret);
1182                         goto out;
1183                 }
1184
1185                 /* Clear suid / sgid if necessary. We do this here
1186                  * instead of later in the write path because
1187                  * remove_suid() calls ->setattr without any hint that
1188                  * we may have already done our cluster locking. Since
1189                  * ocfs2_setattr() *must* take cluster locks to
1190                  * proceeed, this will lead us to recursively lock the
1191                  * inode. There's also the dinode i_size state which
1192                  * can be lost via setattr during extending writes (we
1193                  * set inode->i_size at the end of a write. */
1194                 if (should_remove_suid(dentry)) {
1195                         if (meta_level == 0) {
1196                                 ocfs2_meta_unlock(inode, meta_level);
1197                                 meta_level = 1;
1198                                 continue;
1199                         }
1200
1201                         ret = ocfs2_write_remove_suid(inode);
1202                         if (ret < 0) {
1203                                 mlog_errno(ret);
1204                                 goto out_unlock;
1205                         }
1206                 }
1207
1208                 /* work on a copy of ppos until we're sure that we won't have
1209                  * to recalculate it due to relocking. */
1210                 if (appending) {
1211                         saved_pos = i_size_read(inode);
1212                         mlog(0, "O_APPEND: inode->i_size=%llu\n", saved_pos);
1213                 } else {
1214                         saved_pos = *ppos;
1215                 }
1216
1217                 if (ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb))) {
1218                         loff_t end = saved_pos + count;
1219
1220                         /*
1221                          * Skip the O_DIRECT checks if we don't need
1222                          * them.
1223                          */
1224                         if (!direct_io || !(*direct_io))
1225                                 break;
1226
1227                         /*
1228                          * Allowing concurrent direct writes means
1229                          * i_size changes wouldn't be synchronized, so
1230                          * one node could wind up truncating another
1231                          * nodes writes.
1232                          */
1233                         if (end > i_size_read(inode)) {
1234                                 *direct_io = 0;
1235                                 break;
1236                         }
1237
1238                         /*
1239                          * We don't fill holes during direct io, so
1240                          * check for them here. If any are found, the
1241                          * caller will have to retake some cluster
1242                          * locks and initiate the io as buffered.
1243                          */
1244                         ret = ocfs2_check_range_for_holes(inode, saved_pos,
1245                                                           count);
1246                         if (ret == 1) {
1247                                 *direct_io = 0;
1248                                 ret = 0;
1249                         } else if (ret < 0)
1250                                 mlog_errno(ret);
1251                         break;
1252                 }
1253
1254                 /*
1255                  * The rest of this loop is concerned with legacy file
1256                  * systems which don't support sparse files.
1257                  */
1258
1259                 newsize = count + saved_pos;
1260
1261                 mlog(0, "pos=%lld newsize=%lld cursize=%lld\n",
1262                      (long long) saved_pos, (long long) newsize,
1263                      (long long) i_size_read(inode));
1264
1265                 /* No need for a higher level metadata lock if we're
1266                  * never going past i_size. */
1267                 if (newsize <= i_size_read(inode))
1268                         break;
1269
1270                 if (meta_level == 0) {
1271                         ocfs2_meta_unlock(inode, meta_level);
1272                         meta_level = 1;
1273                         continue;
1274                 }
1275
1276                 spin_lock(&OCFS2_I(inode)->ip_lock);
1277                 clusters = ocfs2_clusters_for_bytes(inode->i_sb, newsize) -
1278                         OCFS2_I(inode)->ip_clusters;
1279                 spin_unlock(&OCFS2_I(inode)->ip_lock);
1280
1281                 mlog(0, "Writing at EOF, may need more allocation: "
1282                      "i_size = %lld, newsize = %lld, need %u clusters\n",
1283                      (long long) i_size_read(inode), (long long) newsize,
1284                      clusters);
1285
1286                 /* We only want to continue the rest of this loop if
1287                  * our extend will actually require more
1288                  * allocation. */
1289                 if (!clusters)
1290                         break;
1291
1292                 ret = ocfs2_extend_file(inode, NULL, newsize, count);
1293                 if (ret < 0) {
1294                         if (ret != -ENOSPC)
1295                                 mlog_errno(ret);
1296                         goto out_unlock;
1297                 }
1298                 break;
1299         }
1300
1301         if (appending)
1302                 *ppos = saved_pos;
1303
1304 out_unlock:
1305         ocfs2_meta_unlock(inode, meta_level);
1306
1307 out:
1308         return ret;
1309 }
1310
1311 static inline void
1312 ocfs2_set_next_iovec(const struct iovec **iovp, size_t *basep, size_t bytes)
1313 {
1314         const struct iovec *iov = *iovp;
1315         size_t base = *basep;
1316
1317         do {
1318                 int copy = min(bytes, iov->iov_len - base);
1319
1320                 bytes -= copy;
1321                 base += copy;
1322                 if (iov->iov_len == base) {
1323                         iov++;
1324                         base = 0;
1325                 }
1326         } while (bytes);
1327         *iovp = iov;
1328         *basep = base;
1329 }
1330
1331 static struct page * ocfs2_get_write_source(struct ocfs2_buffered_write_priv *bp,
1332                                             const struct iovec *cur_iov,
1333                                             size_t iov_offset)
1334 {
1335         int ret;
1336         char *buf;
1337         struct page *src_page = NULL;
1338
1339         buf = cur_iov->iov_base + iov_offset;
1340
1341         if (!segment_eq(get_fs(), KERNEL_DS)) {
1342                 /*
1343                  * Pull in the user page. We want to do this outside
1344                  * of the meta data locks in order to preserve locking
1345                  * order in case of page fault.
1346                  */
1347                 ret = get_user_pages(current, current->mm,
1348                                      (unsigned long)buf & PAGE_CACHE_MASK, 1,
1349                                      0, 0, &src_page, NULL);
1350                 if (ret == 1)
1351                         bp->b_src_buf = kmap(src_page);
1352                 else
1353                         src_page = ERR_PTR(-EFAULT);
1354         } else {
1355                 bp->b_src_buf = buf;
1356         }
1357
1358         return src_page;
1359 }
1360
1361 static void ocfs2_put_write_source(struct ocfs2_buffered_write_priv *bp,
1362                                    struct page *page)
1363 {
1364         if (page) {
1365                 kunmap(page);
1366                 page_cache_release(page);
1367         }
1368 }
1369
1370 static ssize_t ocfs2_file_buffered_write(struct file *file, loff_t *ppos,
1371                                          const struct iovec *iov,
1372                                          unsigned long nr_segs,
1373                                          size_t count,
1374                                          ssize_t o_direct_written)
1375 {
1376         int ret = 0;
1377         ssize_t copied, total = 0;
1378         size_t iov_offset = 0;
1379         const struct iovec *cur_iov = iov;
1380         struct ocfs2_buffered_write_priv bp;
1381         struct page *page;
1382
1383         /*
1384          * handle partial DIO write.  Adjust cur_iov if needed.
1385          */
1386         ocfs2_set_next_iovec(&cur_iov, &iov_offset, o_direct_written);
1387
1388         do {
1389                 bp.b_cur_off = iov_offset;
1390                 bp.b_cur_iov = cur_iov;
1391
1392                 page = ocfs2_get_write_source(&bp, cur_iov, iov_offset);
1393                 if (IS_ERR(page)) {
1394                         ret = PTR_ERR(page);
1395                         goto out;
1396                 }
1397
1398                 copied = ocfs2_buffered_write_cluster(file, *ppos, count,
1399                                                       ocfs2_map_and_write_user_data,
1400                                                       &bp);
1401
1402                 ocfs2_put_write_source(&bp, page);
1403
1404                 if (copied < 0) {
1405                         mlog_errno(copied);
1406                         ret = copied;
1407                         goto out;
1408                 }
1409
1410                 total += copied;
1411                 *ppos = *ppos + copied;
1412                 count -= copied;
1413
1414                 ocfs2_set_next_iovec(&cur_iov, &iov_offset, copied);
1415         } while(count);
1416
1417 out:
1418         return total ? total : ret;
1419 }
1420
1421 static int ocfs2_check_iovec(const struct iovec *iov, size_t *counted,
1422                              unsigned long *nr_segs)
1423 {
1424         size_t ocount;          /* original count */
1425         unsigned long seg;
1426
1427         ocount = 0;
1428         for (seg = 0; seg < *nr_segs; seg++) {
1429                 const struct iovec *iv = &iov[seg];
1430
1431                 /*
1432                  * If any segment has a negative length, or the cumulative
1433                  * length ever wraps negative then return -EINVAL.
1434                  */
1435                 ocount += iv->iov_len;
1436                 if (unlikely((ssize_t)(ocount|iv->iov_len) < 0))
1437                         return -EINVAL;
1438                 if (access_ok(VERIFY_READ, iv->iov_base, iv->iov_len))
1439                         continue;
1440                 if (seg == 0)
1441                         return -EFAULT;
1442                 *nr_segs = seg;
1443                 ocount -= iv->iov_len;  /* This segment is no good */
1444                 break;
1445         }
1446
1447         *counted = ocount;
1448         return 0;
1449 }
1450
1451 static ssize_t ocfs2_file_aio_write(struct kiocb *iocb,
1452                                     const struct iovec *iov,
1453                                     unsigned long nr_segs,
1454                                     loff_t pos)
1455 {
1456         int ret, direct_io, appending, rw_level, have_alloc_sem  = 0;
1457         int can_do_direct, sync = 0;
1458         ssize_t written = 0;
1459         size_t ocount;          /* original count */
1460         size_t count;           /* after file limit checks */
1461         loff_t *ppos = &iocb->ki_pos;
1462         struct file *file = iocb->ki_filp;
1463         struct inode *inode = file->f_path.dentry->d_inode;
1464
1465         mlog_entry("(0x%p, %u, '%.*s')\n", file,
1466                    (unsigned int)nr_segs,
1467                    file->f_path.dentry->d_name.len,
1468                    file->f_path.dentry->d_name.name);
1469
1470         if (iocb->ki_left == 0)
1471                 return 0;
1472
1473         ret = ocfs2_check_iovec(iov, &ocount, &nr_segs);
1474         if (ret)
1475                 return ret;
1476
1477         count = ocount;
1478
1479         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
1480
1481         appending = file->f_flags & O_APPEND ? 1 : 0;
1482         direct_io = file->f_flags & O_DIRECT ? 1 : 0;
1483
1484         mutex_lock(&inode->i_mutex);
1485
1486 relock:
1487         /* to match setattr's i_mutex -> i_alloc_sem -> rw_lock ordering */
1488         if (direct_io) {
1489                 down_read(&inode->i_alloc_sem);
1490                 have_alloc_sem = 1;
1491         }
1492
1493         /* concurrent O_DIRECT writes are allowed */
1494         rw_level = !direct_io;
1495         ret = ocfs2_rw_lock(inode, rw_level);
1496         if (ret < 0) {
1497                 mlog_errno(ret);
1498                 goto out_sems;
1499         }
1500
1501         can_do_direct = direct_io;
1502         ret = ocfs2_prepare_inode_for_write(file->f_path.dentry, ppos,
1503                                             iocb->ki_left, appending,
1504                                             &can_do_direct);
1505         if (ret < 0) {
1506                 mlog_errno(ret);
1507                 goto out;
1508         }
1509
1510         /*
1511          * We can't complete the direct I/O as requested, fall back to
1512          * buffered I/O.
1513          */
1514         if (direct_io && !can_do_direct) {
1515                 ocfs2_rw_unlock(inode, rw_level);
1516                 up_read(&inode->i_alloc_sem);
1517
1518                 have_alloc_sem = 0;
1519                 rw_level = -1;
1520
1521                 direct_io = 0;
1522                 sync = 1;
1523                 goto relock;
1524         }
1525
1526         if (!sync && ((file->f_flags & O_SYNC) || IS_SYNC(inode)))
1527                 sync = 1;
1528
1529         /*
1530          * XXX: Is it ok to execute these checks a second time?
1531          */
1532         ret = generic_write_checks(file, ppos, &count, S_ISBLK(inode->i_mode));
1533         if (ret)
1534                 goto out;
1535
1536         /*
1537          * Set pos so that sync_page_range_nolock() below understands
1538          * where to start from. We might've moved it around via the
1539          * calls above. The range we want to actually sync starts from
1540          * *ppos here.
1541          *
1542          */
1543         pos = *ppos;
1544
1545         /* communicate with ocfs2_dio_end_io */
1546         ocfs2_iocb_set_rw_locked(iocb, rw_level);
1547
1548         if (direct_io) {
1549                 written = generic_file_direct_write(iocb, iov, &nr_segs, *ppos,
1550                                                     ppos, count, ocount);
1551                 if (written < 0) {
1552                         ret = written;
1553                         goto out_dio;
1554                 }
1555         } else {
1556                 written = ocfs2_file_buffered_write(file, ppos, iov, nr_segs,
1557                                                     count, written);
1558                 if (written < 0) {
1559                         ret = written;
1560                         if (ret != -EFAULT || ret != -ENOSPC)
1561                                 mlog_errno(ret);
1562                         goto out;
1563                 }
1564         }
1565
1566 out_dio:
1567         /* buffered aio wouldn't have proper lock coverage today */
1568         BUG_ON(ret == -EIOCBQUEUED && !(file->f_flags & O_DIRECT));
1569
1570         /*
1571          * deep in g_f_a_w_n()->ocfs2_direct_IO we pass in a ocfs2_dio_end_io
1572          * function pointer which is called when o_direct io completes so that
1573          * it can unlock our rw lock.  (it's the clustered equivalent of
1574          * i_alloc_sem; protects truncate from racing with pending ios).
1575          * Unfortunately there are error cases which call end_io and others
1576          * that don't.  so we don't have to unlock the rw_lock if either an
1577          * async dio is going to do it in the future or an end_io after an
1578          * error has already done it.
1579          */
1580         if (ret == -EIOCBQUEUED || !ocfs2_iocb_is_rw_locked(iocb)) {
1581                 rw_level = -1;
1582                 have_alloc_sem = 0;
1583         }
1584
1585 out:
1586         if (rw_level != -1)
1587                 ocfs2_rw_unlock(inode, rw_level);
1588
1589 out_sems:
1590         if (have_alloc_sem)
1591                 up_read(&inode->i_alloc_sem);
1592
1593         if (written > 0 && sync) {
1594                 ssize_t err;
1595
1596                 err = sync_page_range_nolock(inode, file->f_mapping, pos, count);
1597                 if (err < 0)
1598                         written = err;
1599         }
1600
1601         mutex_unlock(&inode->i_mutex);
1602
1603         mlog_exit(ret);
1604         return written ? written : ret;
1605 }
1606
1607 static int ocfs2_splice_write_actor(struct pipe_inode_info *pipe,
1608                                     struct pipe_buffer *buf,
1609                                     struct splice_desc *sd)
1610 {
1611         int ret, count, total = 0;
1612         ssize_t copied = 0;
1613         struct ocfs2_splice_write_priv sp;
1614
1615         ret = buf->ops->pin(pipe, buf);
1616         if (ret)
1617                 goto out;
1618
1619         sp.s_sd = sd;
1620         sp.s_buf = buf;
1621         sp.s_pipe = pipe;
1622         sp.s_offset = sd->pos & ~PAGE_CACHE_MASK;
1623         sp.s_buf_offset = buf->offset;
1624
1625         count = sd->len;
1626         if (count + sp.s_offset > PAGE_CACHE_SIZE)
1627                 count = PAGE_CACHE_SIZE - sp.s_offset;
1628
1629         do {
1630                 /*
1631                  * splice wants us to copy up to one page at a
1632                  * time. For pagesize > cluster size, this means we
1633                  * might enter ocfs2_buffered_write_cluster() more
1634                  * than once, so keep track of our progress here.
1635                  */
1636                 copied = ocfs2_buffered_write_cluster(sd->file,
1637                                                       (loff_t)sd->pos + total,
1638                                                       count,
1639                                                       ocfs2_map_and_write_splice_data,
1640                                                       &sp);
1641                 if (copied < 0) {
1642                         mlog_errno(copied);
1643                         ret = copied;
1644                         goto out;
1645                 }
1646
1647                 count -= copied;
1648                 sp.s_offset += copied;
1649                 sp.s_buf_offset += copied;
1650                 total += copied;
1651         } while (count);
1652
1653         ret = 0;
1654 out:
1655
1656         return total ? total : ret;
1657 }
1658
1659 static ssize_t __ocfs2_file_splice_write(struct pipe_inode_info *pipe,
1660                                          struct file *out,
1661                                          loff_t *ppos,
1662                                          size_t len,
1663                                          unsigned int flags)
1664 {
1665         int ret, err;
1666         struct address_space *mapping = out->f_mapping;
1667         struct inode *inode = mapping->host;
1668
1669         ret = __splice_from_pipe(pipe, out, ppos, len, flags,
1670                                  ocfs2_splice_write_actor);
1671         if (ret > 0) {
1672                 *ppos += ret;
1673
1674                 if (unlikely((out->f_flags & O_SYNC) || IS_SYNC(inode))) {
1675                         err = generic_osync_inode(inode, mapping,
1676                                                   OSYNC_METADATA|OSYNC_DATA);
1677                         if (err)
1678                                 ret = err;
1679                 }
1680         }
1681
1682         return ret;
1683 }
1684
1685 static ssize_t ocfs2_file_splice_write(struct pipe_inode_info *pipe,
1686                                        struct file *out,
1687                                        loff_t *ppos,
1688                                        size_t len,
1689                                        unsigned int flags)
1690 {
1691         int ret;
1692         struct inode *inode = out->f_path.dentry->d_inode;
1693
1694         mlog_entry("(0x%p, 0x%p, %u, '%.*s')\n", out, pipe,
1695                    (unsigned int)len,
1696                    out->f_path.dentry->d_name.len,
1697                    out->f_path.dentry->d_name.name);
1698
1699         inode_double_lock(inode, pipe->inode);
1700
1701         ret = ocfs2_rw_lock(inode, 1);
1702         if (ret < 0) {
1703                 mlog_errno(ret);
1704                 goto out;
1705         }
1706
1707         ret = ocfs2_prepare_inode_for_write(out->f_path.dentry, ppos, len, 0,
1708                                             NULL);
1709         if (ret < 0) {
1710                 mlog_errno(ret);
1711                 goto out_unlock;
1712         }
1713
1714         /* ok, we're done with i_size and alloc work */
1715         ret = __ocfs2_file_splice_write(pipe, out, ppos, len, flags);
1716
1717 out_unlock:
1718         ocfs2_rw_unlock(inode, 1);
1719 out:
1720         inode_double_unlock(inode, pipe->inode);
1721
1722         mlog_exit(ret);
1723         return ret;
1724 }
1725
1726 static ssize_t ocfs2_file_splice_read(struct file *in,
1727                                       loff_t *ppos,
1728                                       struct pipe_inode_info *pipe,
1729                                       size_t len,
1730                                       unsigned int flags)
1731 {
1732         int ret = 0;
1733         struct inode *inode = in->f_path.dentry->d_inode;
1734
1735         mlog_entry("(0x%p, 0x%p, %u, '%.*s')\n", in, pipe,
1736                    (unsigned int)len,
1737                    in->f_path.dentry->d_name.len,
1738                    in->f_path.dentry->d_name.name);
1739
1740         /*
1741          * See the comment in ocfs2_file_aio_read()
1742          */
1743         ret = ocfs2_meta_lock(inode, NULL, 0);
1744         if (ret < 0) {
1745                 mlog_errno(ret);
1746                 goto bail;
1747         }
1748         ocfs2_meta_unlock(inode, 0);
1749
1750         ret = generic_file_splice_read(in, ppos, pipe, len, flags);
1751
1752 bail:
1753         mlog_exit(ret);
1754         return ret;
1755 }
1756
1757 static ssize_t ocfs2_file_aio_read(struct kiocb *iocb,
1758                                    const struct iovec *iov,
1759                                    unsigned long nr_segs,
1760                                    loff_t pos)
1761 {
1762         int ret = 0, rw_level = -1, have_alloc_sem = 0, lock_level = 0;
1763         struct file *filp = iocb->ki_filp;
1764         struct inode *inode = filp->f_path.dentry->d_inode;
1765
1766         mlog_entry("(0x%p, %u, '%.*s')\n", filp,
1767                    (unsigned int)nr_segs,
1768                    filp->f_path.dentry->d_name.len,
1769                    filp->f_path.dentry->d_name.name);
1770
1771         if (!inode) {
1772                 ret = -EINVAL;
1773                 mlog_errno(ret);
1774                 goto bail;
1775         }
1776
1777         /*
1778          * buffered reads protect themselves in ->readpage().  O_DIRECT reads
1779          * need locks to protect pending reads from racing with truncate.
1780          */
1781         if (filp->f_flags & O_DIRECT) {
1782                 down_read(&inode->i_alloc_sem);
1783                 have_alloc_sem = 1;
1784
1785                 ret = ocfs2_rw_lock(inode, 0);
1786                 if (ret < 0) {
1787                         mlog_errno(ret);
1788                         goto bail;
1789                 }
1790                 rw_level = 0;
1791                 /* communicate with ocfs2_dio_end_io */
1792                 ocfs2_iocb_set_rw_locked(iocb, rw_level);
1793         }
1794
1795         /*
1796          * We're fine letting folks race truncates and extending
1797          * writes with read across the cluster, just like they can
1798          * locally. Hence no rw_lock during read.
1799          *
1800          * Take and drop the meta data lock to update inode fields
1801          * like i_size. This allows the checks down below
1802          * generic_file_aio_read() a chance of actually working.
1803          */
1804         ret = ocfs2_meta_lock_atime(inode, filp->f_vfsmnt, &lock_level);
1805         if (ret < 0) {
1806                 mlog_errno(ret);
1807                 goto bail;
1808         }
1809         ocfs2_meta_unlock(inode, lock_level);
1810
1811         ret = generic_file_aio_read(iocb, iov, nr_segs, iocb->ki_pos);
1812         if (ret == -EINVAL)
1813                 mlog(ML_ERROR, "generic_file_aio_read returned -EINVAL\n");
1814
1815         /* buffered aio wouldn't have proper lock coverage today */
1816         BUG_ON(ret == -EIOCBQUEUED && !(filp->f_flags & O_DIRECT));
1817
1818         /* see ocfs2_file_aio_write */
1819         if (ret == -EIOCBQUEUED || !ocfs2_iocb_is_rw_locked(iocb)) {
1820                 rw_level = -1;
1821                 have_alloc_sem = 0;
1822         }
1823
1824 bail:
1825         if (have_alloc_sem)
1826                 up_read(&inode->i_alloc_sem);
1827         if (rw_level != -1)
1828                 ocfs2_rw_unlock(inode, rw_level);
1829         mlog_exit(ret);
1830
1831         return ret;
1832 }
1833
1834 const struct inode_operations ocfs2_file_iops = {
1835         .setattr        = ocfs2_setattr,
1836         .getattr        = ocfs2_getattr,
1837         .permission     = ocfs2_permission,
1838 };
1839
1840 const struct inode_operations ocfs2_special_file_iops = {
1841         .setattr        = ocfs2_setattr,
1842         .getattr        = ocfs2_getattr,
1843         .permission     = ocfs2_permission,
1844 };
1845
1846 const struct file_operations ocfs2_fops = {
1847         .read           = do_sync_read,
1848         .write          = do_sync_write,
1849         .sendfile       = generic_file_sendfile,
1850         .mmap           = ocfs2_mmap,
1851         .fsync          = ocfs2_sync_file,
1852         .release        = ocfs2_file_release,
1853         .open           = ocfs2_file_open,
1854         .aio_read       = ocfs2_file_aio_read,
1855         .aio_write      = ocfs2_file_aio_write,
1856         .ioctl          = ocfs2_ioctl,
1857 #ifdef CONFIG_COMPAT
1858         .compat_ioctl   = ocfs2_compat_ioctl,
1859 #endif
1860         .splice_read    = ocfs2_file_splice_read,
1861         .splice_write   = ocfs2_file_splice_write,
1862 };
1863
1864 const struct file_operations ocfs2_dops = {
1865         .read           = generic_read_dir,
1866         .readdir        = ocfs2_readdir,
1867         .fsync          = ocfs2_sync_file,
1868         .ioctl          = ocfs2_ioctl,
1869 #ifdef CONFIG_COMPAT
1870         .compat_ioctl   = ocfs2_compat_ioctl,
1871 #endif
1872 };