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
  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         u64 cluster_bytes;
 267
 268         mlog_entry_void();
 269
 270         /* TODO: This needs to actually orphan the inode in this
 271          * transaction. */
 272
 273         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
 274         if (IS_ERR(handle)) {
 275                 status = PTR_ERR(handle);
 276                 mlog_errno(status);
 277                 goto out;
 278         }
 279
 280         status = ocfs2_journal_access(handle, inode, fe_bh,
 281                                       OCFS2_JOURNAL_ACCESS_WRITE);
 282         if (status < 0) {
 283                 mlog_errno(status);
 284                 goto out_commit;
 285         }
 286
 287         /*
 288          * Do this before setting i_size.
 289          */
 290         cluster_bytes = ocfs2_align_bytes_to_clusters(inode->i_sb, new_i_size);
 291         status = ocfs2_zero_range_for_truncate(inode, handle, new_i_size,
 292                                                cluster_bytes);
 293         if (status) {
 294                 mlog_errno(status);
 295                 goto out_commit;
 296         }
 297
 298         i_size_write(inode, new_i_size);
 299         inode->i_blocks = ocfs2_align_bytes_to_sectors(new_i_size);
 300         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
 301
 302         di = (struct ocfs2_dinode *) fe_bh->b_data;
 303         di->i_size = cpu_to_le64(new_i_size);
 304         di->i_ctime = di->i_mtime = cpu_to_le64(inode->i_ctime.tv_sec);
 305         di->i_ctime_nsec = di->i_mtime_nsec = cpu_to_le32(inode->i_ctime.tv_nsec);
 306
 307         status = ocfs2_journal_dirty(handle, fe_bh);
 308         if (status < 0)
 309                 mlog_errno(status);
 310
 311 out_commit:
 312         ocfs2_commit_trans(osb, handle);
 313 out:
 314
 315         mlog_exit(status);
 316         return status;
 317 }
 318
 319 static int ocfs2_truncate_file(struct inode *inode,
 320                                struct buffer_head *di_bh,
 321                                u64 new_i_size)
 322 {
 323         int status = 0;
 324         struct ocfs2_dinode *fe = NULL;
 325         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 326         struct ocfs2_truncate_context *tc = NULL;
 327
 328         mlog_entry("(inode = %llu, new_i_size = %llu\n",
 329                    (unsigned long long)OCFS2_I(inode)->ip_blkno,
 330                    (unsigned long long)new_i_size);
 331
 332         fe = (struct ocfs2_dinode *) di_bh->b_data;
 333         if (!OCFS2_IS_VALID_DINODE(fe)) {
 334                 OCFS2_RO_ON_INVALID_DINODE(inode->i_sb, fe);
 335                 status = -EIO;
 336                 goto bail;
 337         }
 338
 339         mlog_bug_on_msg(le64_to_cpu(fe->i_size) != i_size_read(inode),
 340                         "Inode %llu, inode i_size = %lld != di "
 341                         "i_size = %llu, i_flags = 0x%x\n",
 342                         (unsigned long long)OCFS2_I(inode)->ip_blkno,
 343                         i_size_read(inode),
 344                         (unsigned long long)le64_to_cpu(fe->i_size),
 345                         le32_to_cpu(fe->i_flags));
 346
 347         if (new_i_size > le64_to_cpu(fe->i_size)) {
 348                 mlog(0, "asked to truncate file with size (%llu) to size (%llu)!\n",
 349                      (unsigned long long)le64_to_cpu(fe->i_size),
 350                      (unsigned long long)new_i_size);
 351                 status = -EINVAL;
 352                 mlog_errno(status);
 353                 goto bail;
 354         }
 355
 356         mlog(0, "inode %llu, i_size = %llu, new_i_size = %llu\n",
 357              (unsigned long long)le64_to_cpu(fe->i_blkno),
 358              (unsigned long long)le64_to_cpu(fe->i_size),
 359              (unsigned long long)new_i_size);
 360
 361         /* lets handle the simple truncate cases before doing any more
 362          * cluster locking. */
 363         if (new_i_size == le64_to_cpu(fe->i_size))
 364                 goto bail;
 365
 366         down_write(&OCFS2_I(inode)->ip_alloc_sem);
 367
 368         /* This forces other nodes to sync and drop their pages. Do
 369          * this even if we have a truncate without allocation change -
 370          * ocfs2 cluster sizes can be much greater than page size, so
 371          * we have to truncate them anyway.  */
 372         status = ocfs2_data_lock(inode, 1);
 373         if (status < 0) {
 374                 up_write(&OCFS2_I(inode)->ip_alloc_sem);
 375
 376                 mlog_errno(status);
 377                 goto bail;
 378         }
 379
 380         unmap_mapping_range(inode->i_mapping, new_i_size + PAGE_SIZE - 1, 0, 1);
 381         truncate_inode_pages(inode->i_mapping, new_i_size);
 382
 383         /* alright, we're going to need to do a full blown alloc size
 384          * change. Orphan the inode so that recovery can complete the
 385          * truncate if necessary. This does the task of marking
 386          * i_size. */
 387         status = ocfs2_orphan_for_truncate(osb, inode, di_bh, new_i_size);
 388         if (status < 0) {
 389                 mlog_errno(status);
 390                 goto bail_unlock_data;
 391         }
 392
 393         status = ocfs2_prepare_truncate(osb, inode, di_bh, &tc);
 394         if (status < 0) {
 395                 mlog_errno(status);
 396                 goto bail_unlock_data;
 397         }
 398
 399         status = ocfs2_commit_truncate(osb, inode, di_bh, tc);
 400         if (status < 0) {
 401                 mlog_errno(status);
 402                 goto bail_unlock_data;
 403         }
 404
 405         /* TODO: orphan dir cleanup here. */
 406 bail_unlock_data:
 407         ocfs2_data_unlock(inode, 1);
 408
 409         up_write(&OCFS2_I(inode)->ip_alloc_sem);
 410
 411 bail:
 412
 413         mlog_exit(status);
 414         return status;
 415 }
 416
 417 /*
 418  * extend allocation only here.
 419  * we'll update all the disk stuff, and oip->alloc_size
 420  *
 421  * expect stuff to be locked, a transaction started and enough data /
 422  * metadata reservations in the contexts.
 423  *
 424  * Will return -EAGAIN, and a reason if a restart is needed.
 425  * If passed in, *reason will always be set, even in error.
 426  */
 427 int ocfs2_do_extend_allocation(struct ocfs2_super *osb,
 428                                struct inode *inode,
 429                                u32 *logical_offset,
 430                                u32 clusters_to_add,
 431                                int mark_unwritten,
 432                                struct buffer_head *fe_bh,
 433                                handle_t *handle,
 434                                struct ocfs2_alloc_context *data_ac,
 435                                struct ocfs2_alloc_context *meta_ac,
 436                                enum ocfs2_alloc_restarted *reason_ret)
 437 {
 438         int status = 0;
 439         int free_extents;
 440         struct ocfs2_dinode *fe = (struct ocfs2_dinode *) fe_bh->b_data;
 441         enum ocfs2_alloc_restarted reason = RESTART_NONE;
 442         u32 bit_off, num_bits;
 443         u64 block;
 444         u8 flags = 0;
 445
 446         BUG_ON(!clusters_to_add);
 447
 448         if (mark_unwritten)
 449                 flags = OCFS2_EXT_UNWRITTEN;
 450
 451         free_extents = ocfs2_num_free_extents(osb, inode, fe);
 452         if (free_extents < 0) {
 453                 status = free_extents;
 454                 mlog_errno(status);
 455                 goto leave;
 456         }
 457
 458         /* there are two cases which could cause us to EAGAIN in the
 459          * we-need-more-metadata case:
 460          * 1) we haven't reserved *any*
 461          * 2) we are so fragmented, we've needed to add metadata too
 462          *    many times. */
 463         if (!free_extents && !meta_ac) {
 464                 mlog(0, "we haven't reserved any metadata!\n");
 465                 status = -EAGAIN;
 466                 reason = RESTART_META;
 467                 goto leave;
 468         } else if ((!free_extents)
 469                    && (ocfs2_alloc_context_bits_left(meta_ac)
 470                        < ocfs2_extend_meta_needed(fe))) {
 471                 mlog(0, "filesystem is really fragmented...\n");
 472                 status = -EAGAIN;
 473                 reason = RESTART_META;
 474                 goto leave;
 475         }
 476
 477         status = ocfs2_claim_clusters(osb, handle, data_ac, 1,
 478                                       &bit_off, &num_bits);
 479         if (status < 0) {
 480                 if (status != -ENOSPC)
 481                         mlog_errno(status);
 482                 goto leave;
 483         }
 484
 485         BUG_ON(num_bits > clusters_to_add);
 486
 487         /* reserve our write early -- insert_extent may update the inode */
 488         status = ocfs2_journal_access(handle, inode, fe_bh,
 489                                       OCFS2_JOURNAL_ACCESS_WRITE);
 490         if (status < 0) {
 491                 mlog_errno(status);
 492                 goto leave;
 493         }
 494
 495         block = ocfs2_clusters_to_blocks(osb->sb, bit_off);
 496         mlog(0, "Allocating %u clusters at block %u for inode %llu\n",
 497              num_bits, bit_off, (unsigned long long)OCFS2_I(inode)->ip_blkno);
 498         status = ocfs2_insert_extent(osb, handle, inode, fe_bh,
 499                                      *logical_offset, block, num_bits,
 500                                      flags, meta_ac);
 501         if (status < 0) {
 502                 mlog_errno(status);
 503                 goto leave;
 504         }
 505
 506         status = ocfs2_journal_dirty(handle, fe_bh);
 507         if (status < 0) {
 508                 mlog_errno(status);
 509                 goto leave;
 510         }
 511
 512         clusters_to_add -= num_bits;
 513         *logical_offset += num_bits;
 514
 515         if (clusters_to_add) {
 516                 mlog(0, "need to alloc once more, clusters = %u, wanted = "
 517                      "%u\n", fe->i_clusters, clusters_to_add);
 518                 status = -EAGAIN;
 519                 reason = RESTART_TRANS;
 520         }
 521
 522 leave:
 523         mlog_exit(status);
 524         if (reason_ret)
 525                 *reason_ret = reason;
 526         return status;
 527 }
 528
 529 /*
 530  * For a given allocation, determine which allocators will need to be
 531  * accessed, and lock them, reserving the appropriate number of bits.
 532  *
 533  * Sparse file systems call this from ocfs2_write_begin_nolock()
 534  * and ocfs2_allocate_unwritten_extents().
 535  *
 536  * File systems which don't support holes call this from
 537  * ocfs2_extend_allocation().
 538  */
 539 int ocfs2_lock_allocators(struct inode *inode, struct ocfs2_dinode *di,
 540                           u32 clusters_to_add, u32 extents_to_split,
 541                           struct ocfs2_alloc_context **data_ac,
 542                           struct ocfs2_alloc_context **meta_ac)
 543 {
 544         int ret, num_free_extents;
 545         unsigned int max_recs_needed = clusters_to_add + 2 * extents_to_split;
 546         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 547
 548         *meta_ac = NULL;
 549         *data_ac = NULL;
 550
 551         mlog(0, "extend inode %llu, i_size = %lld, di->i_clusters = %u, "
 552              "clusters_to_add = %u, extents_to_split = %u\n",
 553              (unsigned long long)OCFS2_I(inode)->ip_blkno, i_size_read(inode),
 554              le32_to_cpu(di->i_clusters), clusters_to_add, extents_to_split);
 555
 556         num_free_extents = ocfs2_num_free_extents(osb, inode, di);
 557         if (num_free_extents < 0) {
 558                 ret = num_free_extents;
 559                 mlog_errno(ret);
 560                 goto out;
 561         }
 562
 563         /*
 564          * Sparse allocation file systems need to be more conservative
 565          * with reserving room for expansion - the actual allocation
 566          * happens while we've got a journal handle open so re-taking
 567          * a cluster lock (because we ran out of room for another
 568          * extent) will violate ordering rules.
 569          *
 570          * Most of the time we'll only be seeing this 1 cluster at a time
 571          * anyway.
 572          *
 573          * Always lock for any unwritten extents - we might want to
 574          * add blocks during a split.
 575          */
 576         if (!num_free_extents ||
 577             (ocfs2_sparse_alloc(osb) && num_free_extents < max_recs_needed)) {
 578                 ret = ocfs2_reserve_new_metadata(osb, di, meta_ac);
 579                 if (ret < 0) {
 580                         if (ret != -ENOSPC)
 581                                 mlog_errno(ret);
 582                         goto out;
 583                 }
 584         }
 585
 586         ret = ocfs2_reserve_clusters(osb, clusters_to_add, data_ac);
 587         if (ret < 0) {
 588                 if (ret != -ENOSPC)
 589                         mlog_errno(ret);
 590                 goto out;
 591         }
 592
 593 out:
 594         if (ret) {
 595                 if (*meta_ac) {
 596                         ocfs2_free_alloc_context(*meta_ac);
 597                         *meta_ac = NULL;
 598                 }
 599
 600                 /*
 601                  * We cannot have an error and a non null *data_ac.
 602                  */
 603         }
 604
 605         return ret;
 606 }
 607
 608 static int __ocfs2_extend_allocation(struct inode *inode, u32 logical_start,
 609                                      u32 clusters_to_add, int mark_unwritten)
 610 {
 611         int status = 0;
 612         int restart_func = 0;
 613         int credits;
 614         u32 prev_clusters;
 615         struct buffer_head *bh = NULL;
 616         struct ocfs2_dinode *fe = NULL;
 617         handle_t *handle = NULL;
 618         struct ocfs2_alloc_context *data_ac = NULL;
 619         struct ocfs2_alloc_context *meta_ac = NULL;
 620         enum ocfs2_alloc_restarted why;
 621         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 622
 623         mlog_entry("(clusters_to_add = %u)\n", clusters_to_add);
 624
 625         /*
 626          * This function only exists for file systems which don't
 627          * support holes.
 628          */
 629         BUG_ON(mark_unwritten && !ocfs2_sparse_alloc(osb));
 630
 631         status = ocfs2_read_block(osb, OCFS2_I(inode)->ip_blkno, &bh,
 632                                   OCFS2_BH_CACHED, inode);
 633         if (status < 0) {
 634                 mlog_errno(status);
 635                 goto leave;
 636         }
 637
 638         fe = (struct ocfs2_dinode *) bh->b_data;
 639         if (!OCFS2_IS_VALID_DINODE(fe)) {
 640                 OCFS2_RO_ON_INVALID_DINODE(inode->i_sb, fe);
 641                 status = -EIO;
 642                 goto leave;
 643         }
 644
 645 restart_all:
 646         BUG_ON(le32_to_cpu(fe->i_clusters) != OCFS2_I(inode)->ip_clusters);
 647
 648         status = ocfs2_lock_allocators(inode, fe, clusters_to_add, 0, &data_ac,
 649                                        &meta_ac);
 650         if (status) {
 651                 mlog_errno(status);
 652                 goto leave;
 653         }
 654
 655         credits = ocfs2_calc_extend_credits(osb->sb, fe, clusters_to_add);
 656         handle = ocfs2_start_trans(osb, credits);
 657         if (IS_ERR(handle)) {
 658                 status = PTR_ERR(handle);
 659                 handle = NULL;
 660                 mlog_errno(status);
 661                 goto leave;
 662         }
 663
 664 restarted_transaction:
 665         /* reserve a write to the file entry early on - that we if we
 666          * run out of credits in the allocation path, we can still
 667          * update i_size. */
 668         status = ocfs2_journal_access(handle, inode, bh,
 669                                       OCFS2_JOURNAL_ACCESS_WRITE);
 670         if (status < 0) {
 671                 mlog_errno(status);
 672                 goto leave;
 673         }
 674
 675         prev_clusters = OCFS2_I(inode)->ip_clusters;
 676
 677         status = ocfs2_do_extend_allocation(osb,
 678                                             inode,
 679                                             &logical_start,
 680                                             clusters_to_add,
 681                                             mark_unwritten,
 682                                             bh,
 683                                             handle,
 684                                             data_ac,
 685                                             meta_ac,
 686                                             &why);
 687         if ((status < 0) && (status != -EAGAIN)) {
 688                 if (status != -ENOSPC)
 689                         mlog_errno(status);
 690                 goto leave;
 691         }
 692
 693         status = ocfs2_journal_dirty(handle, bh);
 694         if (status < 0) {
 695                 mlog_errno(status);
 696                 goto leave;
 697         }
 698
 699         spin_lock(&OCFS2_I(inode)->ip_lock);
 700         clusters_to_add -= (OCFS2_I(inode)->ip_clusters - prev_clusters);
 701         spin_unlock(&OCFS2_I(inode)->ip_lock);
 702
 703         if (why != RESTART_NONE && clusters_to_add) {
 704                 if (why == RESTART_META) {
 705                         mlog(0, "restarting function.\n");
 706                         restart_func = 1;
 707                 } else {
 708                         BUG_ON(why != RESTART_TRANS);
 709
 710                         mlog(0, "restarting transaction.\n");
 711                         /* TODO: This can be more intelligent. */
 712                         credits = ocfs2_calc_extend_credits(osb->sb,
 713                                                             fe,
 714                                                             clusters_to_add);
 715                         status = ocfs2_extend_trans(handle, credits);
 716                         if (status < 0) {
 717                                 /* handle still has to be committed at
 718                                  * this point. */
 719                                 status = -ENOMEM;
 720                                 mlog_errno(status);
 721                                 goto leave;
 722                         }
 723                         goto restarted_transaction;
 724                 }
 725         }
 726
 727         mlog(0, "fe: i_clusters = %u, i_size=%llu\n",
 728              le32_to_cpu(fe->i_clusters),
 729              (unsigned long long)le64_to_cpu(fe->i_size));
 730         mlog(0, "inode: ip_clusters=%u, i_size=%lld\n",
 731              OCFS2_I(inode)->ip_clusters, i_size_read(inode));
 732
 733 leave:
 734         if (handle) {
 735                 ocfs2_commit_trans(osb, handle);
 736                 handle = NULL;
 737         }
 738         if (data_ac) {
 739                 ocfs2_free_alloc_context(data_ac);
 740                 data_ac = NULL;
 741         }
 742         if (meta_ac) {
 743                 ocfs2_free_alloc_context(meta_ac);
 744                 meta_ac = NULL;
 745         }
 746         if ((!status) && restart_func) {
 747                 restart_func = 0;
 748                 goto restart_all;
 749         }
 750         if (bh) {
 751                 brelse(bh);
 752                 bh = NULL;
 753         }
 754
 755         mlog_exit(status);
 756         return status;
 757 }
 758
 759 static int ocfs2_extend_allocation(struct inode *inode, u32 logical_start,
 760                                    u32 clusters_to_add, int mark_unwritten)
 761 {
 762         int ret;
 763
 764         /*
 765          * The alloc sem blocks peope in read/write from reading our
 766          * allocation until we're done changing it. We depend on
 767          * i_mutex to block other extend/truncate calls while we're
 768          * here.
 769          */
 770         down_write(&OCFS2_I(inode)->ip_alloc_sem);
 771         ret = __ocfs2_extend_allocation(inode, logical_start, clusters_to_add,
 772                                         mark_unwritten);
 773         up_write(&OCFS2_I(inode)->ip_alloc_sem);
 774
 775         return ret;
 776 }
 777
 778 /* Some parts of this taken from generic_cont_expand, which turned out
 779  * to be too fragile to do exactly what we need without us having to
 780  * worry about recursive locking in ->prepare_write() and
 781  * ->commit_write(). */
 782 static int ocfs2_write_zero_page(struct inode *inode,
 783                                  u64 size)
 784 {
 785         struct address_space *mapping = inode->i_mapping;
 786         struct page *page;
 787         unsigned long index;
 788         unsigned int offset;
 789         handle_t *handle = NULL;
 790         int ret;
 791
 792         offset = (size & (PAGE_CACHE_SIZE-1)); /* Within page */
 793         /* ugh.  in prepare/commit_write, if from==to==start of block, we
 794         ** skip the prepare.  make sure we never send an offset for the start
 795         ** of a block
 796         */
 797         if ((offset & (inode->i_sb->s_blocksize - 1)) == 0) {
 798                 offset++;
 799         }
 800         index = size >> PAGE_CACHE_SHIFT;
 801
 802         page = grab_cache_page(mapping, index);
 803         if (!page) {
 804                 ret = -ENOMEM;
 805                 mlog_errno(ret);
 806                 goto out;
 807         }
 808
 809         ret = ocfs2_prepare_write_nolock(inode, page, offset, offset);
 810         if (ret < 0) {
 811                 mlog_errno(ret);
 812                 goto out_unlock;
 813         }
 814
 815         if (ocfs2_should_order_data(inode)) {
 816                 handle = ocfs2_start_walk_page_trans(inode, page, offset,
 817                                                      offset);
 818                 if (IS_ERR(handle)) {
 819                         ret = PTR_ERR(handle);
 820                         handle = NULL;
 821                         goto out_unlock;
 822                 }
 823         }
 824
 825         /* must not update i_size! */
 826         ret = block_commit_write(page, offset, offset);
 827         if (ret < 0)
 828                 mlog_errno(ret);
 829         else
 830                 ret = 0;
 831
 832         if (handle)
 833                 ocfs2_commit_trans(OCFS2_SB(inode->i_sb), handle);
 834 out_unlock:
 835         unlock_page(page);
 836         page_cache_release(page);
 837 out:
 838         return ret;
 839 }
 840
 841 static int ocfs2_zero_extend(struct inode *inode,
 842                              u64 zero_to_size)
 843 {
 844         int ret = 0;
 845         u64 start_off;
 846         struct super_block *sb = inode->i_sb;
 847
 848         start_off = ocfs2_align_bytes_to_blocks(sb, i_size_read(inode));
 849         while (start_off < zero_to_size) {
 850                 ret = ocfs2_write_zero_page(inode, start_off);
 851                 if (ret < 0) {
 852                         mlog_errno(ret);
 853                         goto out;
 854                 }
 855
 856                 start_off += sb->s_blocksize;
 857
 858                 /*
 859                  * Very large extends have the potential to lock up
 860                  * the cpu for extended periods of time.
 861                  */
 862                 cond_resched();
 863         }
 864
 865 out:
 866         return ret;
 867 }
 868
 869 /*
 870  * A tail_to_skip value > 0 indicates that we're being called from
 871  * ocfs2_file_aio_write(). This has the following implications:
 872  *
 873  * - we don't want to update i_size
 874  * - di_bh will be NULL, which is fine because it's only used in the
 875  *   case where we want to update i_size.
 876  * - ocfs2_zero_extend() will then only be filling the hole created
 877  *   between i_size and the start of the write.
 878  */
 879 static int ocfs2_extend_file(struct inode *inode,
 880                              struct buffer_head *di_bh,
 881                              u64 new_i_size,
 882                              size_t tail_to_skip)
 883 {
 884         int ret = 0;
 885         u32 clusters_to_add = 0;
 886
 887         BUG_ON(!tail_to_skip && !di_bh);
 888
 889         /* setattr sometimes calls us like this. */
 890         if (new_i_size == 0)
 891                 goto out;
 892
 893         if (i_size_read(inode) == new_i_size)
 894                 goto out;
 895         BUG_ON(new_i_size < i_size_read(inode));
 896
 897         if (ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb))) {
 898                 BUG_ON(tail_to_skip != 0);
 899                 goto out_update_size;
 900         }
 901
 902         clusters_to_add = ocfs2_clusters_for_bytes(inode->i_sb, new_i_size) -
 903                 OCFS2_I(inode)->ip_clusters;
 904
 905         /*
 906          * protect the pages that ocfs2_zero_extend is going to be
 907          * pulling into the page cache.. we do this before the
 908          * metadata extend so that we don't get into the situation
 909          * where we've extended the metadata but can't get the data
 910          * lock to zero.
 911          */
 912         ret = ocfs2_data_lock(inode, 1);
 913         if (ret < 0) {
 914                 mlog_errno(ret);
 915                 goto out;
 916         }
 917
 918         if (clusters_to_add) {
 919                 ret = ocfs2_extend_allocation(inode,
 920                                               OCFS2_I(inode)->ip_clusters,
 921                                               clusters_to_add, 0);
 922                 if (ret < 0) {
 923                         mlog_errno(ret);
 924                         goto out_unlock;
 925                 }
 926         }
 927
 928         /*
 929          * Call this even if we don't add any clusters to the tree. We
 930          * still need to zero the area between the old i_size and the
 931          * new i_size.
 932          */
 933         ret = ocfs2_zero_extend(inode, (u64)new_i_size - tail_to_skip);
 934         if (ret < 0) {
 935                 mlog_errno(ret);
 936                 goto out_unlock;
 937         }
 938
 939 out_update_size:
 940         if (!tail_to_skip) {
 941                 /* We're being called from ocfs2_setattr() which wants
 942                  * us to update i_size */
 943                 ret = ocfs2_simple_size_update(inode, di_bh, new_i_size);
 944                 if (ret < 0)
 945                         mlog_errno(ret);
 946         }
 947
 948 out_unlock:
 949         if (!ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)))
 950                 ocfs2_data_unlock(inode, 1);
 951
 952 out:
 953         return ret;
 954 }
 955
 956 int ocfs2_setattr(struct dentry *dentry, struct iattr *attr)
 957 {
 958         int status = 0, size_change;
 959         struct inode *inode = dentry->d_inode;
 960         struct super_block *sb = inode->i_sb;
 961         struct ocfs2_super *osb = OCFS2_SB(sb);
 962         struct buffer_head *bh = NULL;
 963         handle_t *handle = NULL;
 964
 965         mlog_entry("(0x%p, '%.*s')\n", dentry,
 966                    dentry->d_name.len, dentry->d_name.name);
 967
 968         if (attr->ia_valid & ATTR_MODE)
 969                 mlog(0, "mode change: %d\n", attr->ia_mode);
 970         if (attr->ia_valid & ATTR_UID)
 971                 mlog(0, "uid change: %d\n", attr->ia_uid);
 972         if (attr->ia_valid & ATTR_GID)
 973                 mlog(0, "gid change: %d\n", attr->ia_gid);
 974         if (attr->ia_valid & ATTR_SIZE)
 975                 mlog(0, "size change...\n");
 976         if (attr->ia_valid & (ATTR_ATIME | ATTR_MTIME | ATTR_CTIME))
 977                 mlog(0, "time change...\n");
 978
 979 #define OCFS2_VALID_ATTRS (ATTR_ATIME | ATTR_MTIME | ATTR_CTIME | ATTR_SIZE \
 980                            | ATTR_GID | ATTR_UID | ATTR_MODE)
 981         if (!(attr->ia_valid & OCFS2_VALID_ATTRS)) {
 982                 mlog(0, "can't handle attrs: 0x%x\n", attr->ia_valid);
 983                 return 0;
 984         }
 985
 986         status = inode_change_ok(inode, attr);
 987         if (status)
 988                 return status;
 989
 990         size_change = S_ISREG(inode->i_mode) && attr->ia_valid & ATTR_SIZE;
 991         if (size_change) {
 992                 status = ocfs2_rw_lock(inode, 1);
 993                 if (status < 0) {
 994                         mlog_errno(status);
 995                         goto bail;
 996                 }
 997         }
 998
 999         status = ocfs2_meta_lock(inode, &bh, 1);
1000         if (status < 0) {
1001                 if (status != -ENOENT)
1002                         mlog_errno(status);
1003                 goto bail_unlock_rw;
1004         }
1005
1006         if (size_change && attr->ia_size != i_size_read(inode)) {
1007                 if (i_size_read(inode) > attr->ia_size)
1008                         status = ocfs2_truncate_file(inode, bh, attr->ia_size);
1009                 else
1010                         status = ocfs2_extend_file(inode, bh, attr->ia_size, 0);
1011                 if (status < 0) {
1012                         if (status != -ENOSPC)
1013                                 mlog_errno(status);
1014                         status = -ENOSPC;
1015                         goto bail_unlock;
1016                 }
1017         }
1018
1019         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1020         if (IS_ERR(handle)) {
1021                 status = PTR_ERR(handle);
1022                 mlog_errno(status);
1023                 goto bail_unlock;
1024         }
1025
1026         /*
1027          * This will intentionally not wind up calling vmtruncate(),
1028          * since all the work for a size change has been done above.
1029          * Otherwise, we could get into problems with truncate as
1030          * ip_alloc_sem is used there to protect against i_size
1031          * changes.
1032          */
1033         status = inode_setattr(inode, attr);
1034         if (status < 0) {
1035                 mlog_errno(status);
1036                 goto bail_commit;
1037         }
1038
1039         status = ocfs2_mark_inode_dirty(handle, inode, bh);
1040         if (status < 0)
1041                 mlog_errno(status);
1042
1043 bail_commit:
1044         ocfs2_commit_trans(osb, handle);
1045 bail_unlock:
1046         ocfs2_meta_unlock(inode, 1);
1047 bail_unlock_rw:
1048         if (size_change)
1049                 ocfs2_rw_unlock(inode, 1);
1050 bail:
1051         if (bh)
1052                 brelse(bh);
1053
1054         mlog_exit(status);
1055         return status;
1056 }
1057
1058 int ocfs2_getattr(struct vfsmount *mnt,
1059                   struct dentry *dentry,
1060                   struct kstat *stat)
1061 {
1062         struct inode *inode = dentry->d_inode;
1063         struct super_block *sb = dentry->d_inode->i_sb;
1064         struct ocfs2_super *osb = sb->s_fs_info;
1065         int err;
1066
1067         mlog_entry_void();
1068
1069         err = ocfs2_inode_revalidate(dentry);
1070         if (err) {
1071                 if (err != -ENOENT)
1072                         mlog_errno(err);
1073                 goto bail;
1074         }
1075
1076         generic_fillattr(inode, stat);
1077
1078         /* We set the blksize from the cluster size for performance */
1079         stat->blksize = osb->s_clustersize;
1080
1081 bail:
1082         mlog_exit(err);
1083
1084         return err;
1085 }
1086
1087 int ocfs2_permission(struct inode *inode, int mask, struct nameidata *nd)
1088 {
1089         int ret;
1090
1091         mlog_entry_void();
1092
1093         ret = ocfs2_meta_lock(inode, NULL, 0);
1094         if (ret) {
1095                 if (ret != -ENOENT)
1096                         mlog_errno(ret);
1097                 goto out;
1098         }
1099
1100         ret = generic_permission(inode, mask, NULL);
1101
1102         ocfs2_meta_unlock(inode, 0);
1103 out:
1104         mlog_exit(ret);
1105         return ret;
1106 }
1107
1108 static int ocfs2_write_remove_suid(struct inode *inode)
1109 {
1110         int ret;
1111         struct buffer_head *bh = NULL;
1112         struct ocfs2_inode_info *oi = OCFS2_I(inode);
1113         handle_t *handle;
1114         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1115         struct ocfs2_dinode *di;
1116
1117         mlog_entry("(Inode %llu, mode 0%o)\n",
1118                    (unsigned long long)oi->ip_blkno, inode->i_mode);
1119
1120         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1121         if (handle == NULL) {
1122                 ret = -ENOMEM;
1123                 mlog_errno(ret);
1124                 goto out;
1125         }
1126
1127         ret = ocfs2_read_block(osb, oi->ip_blkno, &bh, OCFS2_BH_CACHED, inode);
1128         if (ret < 0) {
1129                 mlog_errno(ret);
1130                 goto out_trans;
1131         }
1132
1133         ret = ocfs2_journal_access(handle, inode, bh,
1134                                    OCFS2_JOURNAL_ACCESS_WRITE);
1135         if (ret < 0) {
1136                 mlog_errno(ret);
1137                 goto out_bh;
1138         }
1139
1140         inode->i_mode &= ~S_ISUID;
1141         if ((inode->i_mode & S_ISGID) && (inode->i_mode & S_IXGRP))
1142                 inode->i_mode &= ~S_ISGID;
1143
1144         di = (struct ocfs2_dinode *) bh->b_data;
1145         di->i_mode = cpu_to_le16(inode->i_mode);
1146
1147         ret = ocfs2_journal_dirty(handle, bh);
1148         if (ret < 0)
1149                 mlog_errno(ret);
1150 out_bh:
1151         brelse(bh);
1152 out_trans:
1153         ocfs2_commit_trans(osb, handle);
1154 out:
1155         mlog_exit(ret);
1156         return ret;
1157 }
1158
1159 /*
1160  * Will look for holes and unwritten extents in the range starting at
1161  * pos for count bytes (inclusive).
1162  */
1163 static int ocfs2_check_range_for_holes(struct inode *inode, loff_t pos,
1164                                        size_t count)
1165 {
1166         int ret = 0;
1167         unsigned int extent_flags;
1168         u32 cpos, clusters, extent_len, phys_cpos;
1169         struct super_block *sb = inode->i_sb;
1170
1171         cpos = pos >> OCFS2_SB(sb)->s_clustersize_bits;
1172         clusters = ocfs2_clusters_for_bytes(sb, pos + count) - cpos;
1173
1174         while (clusters) {
1175                 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos, &extent_len,
1176                                          &extent_flags);
1177                 if (ret < 0) {
1178                         mlog_errno(ret);
1179                         goto out;
1180                 }
1181
1182                 if (phys_cpos == 0 || (extent_flags & OCFS2_EXT_UNWRITTEN)) {
1183                         ret = 1;
1184                         break;
1185                 }
1186
1187                 if (extent_len > clusters)
1188                         extent_len = clusters;
1189
1190                 clusters -= extent_len;
1191                 cpos += extent_len;
1192         }
1193 out:
1194         return ret;
1195 }
1196
1197 /*
1198  * Allocate enough extents to cover the region starting at byte offset
1199  * start for len bytes. Existing extents are skipped, any extents
1200  * added are marked as "unwritten".
1201  */
1202 static int ocfs2_allocate_unwritten_extents(struct inode *inode,
1203                                             u64 start, u64 len)
1204 {
1205         int ret;
1206         u32 cpos, phys_cpos, clusters, alloc_size;
1207
1208         /*
1209          * We consider both start and len to be inclusive.
1210          */
1211         cpos = start >> OCFS2_SB(inode->i_sb)->s_clustersize_bits;
1212         clusters = ocfs2_clusters_for_bytes(inode->i_sb, start + len);
1213         clusters -= cpos;
1214
1215         while (clusters) {
1216                 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos,
1217                                          &alloc_size, NULL);
1218                 if (ret) {
1219                         mlog_errno(ret);
1220                         goto out;
1221                 }
1222
1223                 /*
1224                  * Hole or existing extent len can be arbitrary, so
1225                  * cap it to our own allocation request.
1226                  */
1227                 if (alloc_size > clusters)
1228                         alloc_size = clusters;
1229
1230                 if (phys_cpos) {
1231                         /*
1232                          * We already have an allocation at this
1233                          * region so we can safely skip it.
1234                          */
1235                         goto next;
1236                 }
1237
1238                 ret = __ocfs2_extend_allocation(inode, cpos, alloc_size, 1);
1239                 if (ret) {
1240                         if (ret != -ENOSPC)
1241                                 mlog_errno(ret);
1242                         goto out;
1243                 }
1244
1245 next:
1246                 cpos += alloc_size;
1247                 clusters -= alloc_size;
1248         }
1249
1250         ret = 0;
1251 out:
1252         return ret;
1253 }
1254
1255 static int ocfs2_prepare_inode_for_write(struct dentry *dentry,
1256                                          loff_t *ppos,
1257                                          size_t count,
1258                                          int appending,
1259                                          int *direct_io)
1260 {
1261         int ret = 0, meta_level = appending;
1262         struct inode *inode = dentry->d_inode;
1263         u32 clusters;
1264         loff_t newsize, saved_pos;
1265
1266         /*
1267          * We sample i_size under a read level meta lock to see if our write
1268          * is extending the file, if it is we back off and get a write level
1269          * meta lock.
1270          */
1271         for(;;) {
1272                 ret = ocfs2_meta_lock(inode, NULL, meta_level);
1273                 if (ret < 0) {
1274                         meta_level = -1;
1275                         mlog_errno(ret);
1276                         goto out;
1277                 }
1278
1279                 /* Clear suid / sgid if necessary. We do this here
1280                  * instead of later in the write path because
1281                  * remove_suid() calls ->setattr without any hint that
1282                  * we may have already done our cluster locking. Since
1283                  * ocfs2_setattr() *must* take cluster locks to
1284                  * proceeed, this will lead us to recursively lock the
1285                  * inode. There's also the dinode i_size state which
1286                  * can be lost via setattr during extending writes (we
1287                  * set inode->i_size at the end of a write. */
1288                 if (should_remove_suid(dentry)) {
1289                         if (meta_level == 0) {
1290                                 ocfs2_meta_unlock(inode, meta_level);
1291                                 meta_level = 1;
1292                                 continue;
1293                         }
1294
1295                         ret = ocfs2_write_remove_suid(inode);
1296                         if (ret < 0) {
1297                                 mlog_errno(ret);
1298                                 goto out_unlock;
1299                         }
1300                 }
1301
1302                 /* work on a copy of ppos until we're sure that we won't have
1303                  * to recalculate it due to relocking. */
1304                 if (appending) {
1305                         saved_pos = i_size_read(inode);
1306                         mlog(0, "O_APPEND: inode->i_size=%llu\n", saved_pos);
1307                 } else {
1308                         saved_pos = *ppos;
1309                 }
1310
1311                 if (ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb))) {
1312                         loff_t end = saved_pos + count;
1313
1314                         /*
1315                          * Skip the O_DIRECT checks if we don't need
1316                          * them.
1317                          */
1318                         if (!direct_io || !(*direct_io))
1319                                 break;
1320
1321                         /*
1322                          * Allowing concurrent direct writes means
1323                          * i_size changes wouldn't be synchronized, so
1324                          * one node could wind up truncating another
1325                          * nodes writes.
1326                          */
1327                         if (end > i_size_read(inode)) {
1328                                 *direct_io = 0;
1329                                 break;
1330                         }
1331
1332                         /*
1333                          * We don't fill holes during direct io, so
1334                          * check for them here. If any are found, the
1335                          * caller will have to retake some cluster
1336                          * locks and initiate the io as buffered.
1337                          */
1338                         ret = ocfs2_check_range_for_holes(inode, saved_pos,
1339                                                           count);
1340                         if (ret == 1) {
1341                                 *direct_io = 0;
1342                                 ret = 0;
1343                         } else if (ret < 0)
1344                                 mlog_errno(ret);
1345                         break;
1346                 }
1347
1348                 /*
1349                  * The rest of this loop is concerned with legacy file
1350                  * systems which don't support sparse files.
1351                  */
1352
1353                 newsize = count + saved_pos;
1354
1355                 mlog(0, "pos=%lld newsize=%lld cursize=%lld\n",
1356                      (long long) saved_pos, (long long) newsize,
1357                      (long long) i_size_read(inode));
1358
1359                 /* No need for a higher level metadata lock if we're
1360                  * never going past i_size. */
1361                 if (newsize <= i_size_read(inode))
1362                         break;
1363
1364                 if (meta_level == 0) {
1365                         ocfs2_meta_unlock(inode, meta_level);
1366                         meta_level = 1;
1367                         continue;
1368                 }
1369
1370                 spin_lock(&OCFS2_I(inode)->ip_lock);
1371                 clusters = ocfs2_clusters_for_bytes(inode->i_sb, newsize) -
1372                         OCFS2_I(inode)->ip_clusters;
1373                 spin_unlock(&OCFS2_I(inode)->ip_lock);
1374
1375                 mlog(0, "Writing at EOF, may need more allocation: "
1376                      "i_size = %lld, newsize = %lld, need %u clusters\n",
1377                      (long long) i_size_read(inode), (long long) newsize,
1378                      clusters);
1379
1380                 /* We only want to continue the rest of this loop if
1381                  * our extend will actually require more
1382                  * allocation. */
1383                 if (!clusters)
1384                         break;
1385
1386                 ret = ocfs2_extend_file(inode, NULL, newsize, count);
1387                 if (ret < 0) {
1388                         if (ret != -ENOSPC)
1389                                 mlog_errno(ret);
1390                         goto out_unlock;
1391                 }
1392                 break;
1393         }
1394
1395         if (appending)
1396                 *ppos = saved_pos;
1397
1398 out_unlock:
1399         ocfs2_meta_unlock(inode, meta_level);
1400
1401 out:
1402         return ret;
1403 }
1404
1405 static inline void
1406 ocfs2_set_next_iovec(const struct iovec **iovp, size_t *basep, size_t bytes)
1407 {
1408         const struct iovec *iov = *iovp;
1409         size_t base = *basep;
1410
1411         do {
1412                 int copy = min(bytes, iov->iov_len - base);
1413
1414                 bytes -= copy;
1415                 base += copy;
1416                 if (iov->iov_len == base) {
1417                         iov++;
1418                         base = 0;
1419                 }
1420         } while (bytes);
1421         *iovp = iov;
1422         *basep = base;
1423 }
1424
1425 static struct page * ocfs2_get_write_source(char **ret_src_buf,
1426                                             const struct iovec *cur_iov,
1427                                             size_t iov_offset)
1428 {
1429         int ret;
1430         char *buf = cur_iov->iov_base + iov_offset;
1431         struct page *src_page = NULL;
1432         unsigned long off;
1433
1434         off = (unsigned long)(buf) & ~PAGE_CACHE_MASK;
1435
1436         if (!segment_eq(get_fs(), KERNEL_DS)) {
1437                 /*
1438                  * Pull in the user page. We want to do this outside
1439                  * of the meta data locks in order to preserve locking
1440                  * order in case of page fault.
1441                  */
1442                 ret = get_user_pages(current, current->mm,
1443                                      (unsigned long)buf & PAGE_CACHE_MASK, 1,
1444                                      0, 0, &src_page, NULL);
1445                 if (ret == 1)
1446                         *ret_src_buf = kmap(src_page) + off;
1447                 else
1448                         src_page = ERR_PTR(-EFAULT);
1449         } else {
1450                 *ret_src_buf = buf;
1451         }
1452
1453         return src_page;
1454 }
1455
1456 static void ocfs2_put_write_source(struct page *page)
1457 {
1458         if (page) {
1459                 kunmap(page);
1460                 page_cache_release(page);
1461         }
1462 }
1463
1464 static ssize_t ocfs2_file_buffered_write(struct file *file, loff_t *ppos,
1465                                          const struct iovec *iov,
1466                                          unsigned long nr_segs,
1467                                          size_t count,
1468                                          ssize_t o_direct_written)
1469 {
1470         int ret = 0;
1471         ssize_t copied, total = 0;
1472         size_t iov_offset = 0, bytes;
1473         loff_t pos;
1474         const struct iovec *cur_iov = iov;
1475         struct page *user_page, *page;
1476         char *buf, *dst;
1477         void *fsdata;
1478
1479         /*
1480          * handle partial DIO write.  Adjust cur_iov if needed.
1481          */
1482         ocfs2_set_next_iovec(&cur_iov, &iov_offset, o_direct_written);
1483
1484         do {
1485                 pos = *ppos;
1486
1487                 user_page = ocfs2_get_write_source(&buf, cur_iov, iov_offset);
1488                 if (IS_ERR(user_page)) {
1489                         ret = PTR_ERR(user_page);
1490                         goto out;
1491                 }
1492
1493                 /* Stay within our page boundaries */
1494                 bytes = min((PAGE_CACHE_SIZE - ((unsigned long)pos & ~PAGE_CACHE_MASK)),
1495                             (PAGE_CACHE_SIZE - ((unsigned long)buf & ~PAGE_CACHE_MASK)));
1496                 /* Stay within the vector boundary */
1497                 bytes = min_t(size_t, bytes, cur_iov->iov_len - iov_offset);
1498                 /* Stay within count */
1499                 bytes = min(bytes, count);
1500
1501                 page = NULL;
1502                 ret = ocfs2_write_begin(file, file->f_mapping, pos, bytes, 0,
1503                                         &page, &fsdata);
1504                 if (ret) {
1505                         mlog_errno(ret);
1506                         goto out;
1507                 }
1508
1509                 dst = kmap_atomic(page, KM_USER0);
1510                 memcpy(dst + (pos & (PAGE_CACHE_SIZE - 1)), buf, bytes);
1511                 kunmap_atomic(dst, KM_USER0);
1512                 flush_dcache_page(page);
1513                 ocfs2_put_write_source(user_page);
1514
1515                 copied = ocfs2_write_end(file, file->f_mapping, pos, bytes,
1516                                          bytes, page, fsdata);
1517                 if (copied < 0) {
1518                         mlog_errno(copied);
1519                         ret = copied;
1520                         goto out;
1521                 }
1522
1523                 total += copied;
1524                 *ppos = pos + copied;
1525                 count -= copied;
1526
1527                 ocfs2_set_next_iovec(&cur_iov, &iov_offset, copied);
1528         } while(count);
1529
1530 out:
1531         return total ? total : ret;
1532 }
1533
1534 static ssize_t ocfs2_file_aio_write(struct kiocb *iocb,
1535                                     const struct iovec *iov,
1536                                     unsigned long nr_segs,
1537                                     loff_t pos)
1538 {
1539         int ret, direct_io, appending, rw_level, have_alloc_sem  = 0;
1540         int can_do_direct, sync = 0;
1541         ssize_t written = 0;
1542         size_t ocount;          /* original count */
1543         size_t count;           /* after file limit checks */
1544         loff_t *ppos = &iocb->ki_pos;
1545         struct file *file = iocb->ki_filp;
1546         struct inode *inode = file->f_path.dentry->d_inode;
1547
1548         mlog_entry("(0x%p, %u, '%.*s')\n", file,
1549                    (unsigned int)nr_segs,
1550                    file->f_path.dentry->d_name.len,
1551                    file->f_path.dentry->d_name.name);
1552
1553         if (iocb->ki_left == 0)
1554                 return 0;
1555
1556         ret = generic_segment_checks(iov, &nr_segs, &ocount, VERIFY_READ);
1557         if (ret)
1558                 return ret;
1559
1560         count = ocount;
1561
1562         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
1563
1564         appending = file->f_flags & O_APPEND ? 1 : 0;
1565         direct_io = file->f_flags & O_DIRECT ? 1 : 0;
1566
1567         mutex_lock(&inode->i_mutex);
1568
1569 relock:
1570         /* to match setattr's i_mutex -> i_alloc_sem -> rw_lock ordering */
1571         if (direct_io) {
1572                 down_read(&inode->i_alloc_sem);
1573                 have_alloc_sem = 1;
1574         }
1575
1576         /* concurrent O_DIRECT writes are allowed */
1577         rw_level = !direct_io;
1578         ret = ocfs2_rw_lock(inode, rw_level);
1579         if (ret < 0) {
1580                 mlog_errno(ret);
1581                 goto out_sems;
1582         }
1583
1584         can_do_direct = direct_io;
1585         ret = ocfs2_prepare_inode_for_write(file->f_path.dentry, ppos,
1586                                             iocb->ki_left, appending,
1587                                             &can_do_direct);
1588         if (ret < 0) {
1589                 mlog_errno(ret);
1590                 goto out;
1591         }
1592
1593         /*
1594          * We can't complete the direct I/O as requested, fall back to
1595          * buffered I/O.
1596          */
1597         if (direct_io && !can_do_direct) {
1598                 ocfs2_rw_unlock(inode, rw_level);
1599                 up_read(&inode->i_alloc_sem);
1600
1601                 have_alloc_sem = 0;
1602                 rw_level = -1;
1603
1604                 direct_io = 0;
1605                 sync = 1;
1606                 goto relock;
1607         }
1608
1609         if (!sync && ((file->f_flags & O_SYNC) || IS_SYNC(inode)))
1610                 sync = 1;
1611
1612         /*
1613          * XXX: Is it ok to execute these checks a second time?
1614          */
1615         ret = generic_write_checks(file, ppos, &count, S_ISBLK(inode->i_mode));
1616         if (ret)
1617                 goto out;
1618
1619         /*
1620          * Set pos so that sync_page_range_nolock() below understands
1621          * where to start from. We might've moved it around via the
1622          * calls above. The range we want to actually sync starts from
1623          * *ppos here.
1624          *
1625          */
1626         pos = *ppos;
1627
1628         /* communicate with ocfs2_dio_end_io */
1629         ocfs2_iocb_set_rw_locked(iocb, rw_level);
1630
1631         if (direct_io) {
1632                 written = generic_file_direct_write(iocb, iov, &nr_segs, *ppos,
1633                                                     ppos, count, ocount);
1634                 if (written < 0) {
1635                         ret = written;
1636                         goto out_dio;
1637                 }
1638         } else {
1639                 written = ocfs2_file_buffered_write(file, ppos, iov, nr_segs,
1640                                                     count, written);
1641                 if (written < 0) {
1642                         ret = written;
1643                         if (ret != -EFAULT || ret != -ENOSPC)
1644                                 mlog_errno(ret);
1645                         goto out;
1646                 }
1647         }
1648
1649 out_dio:
1650         /* buffered aio wouldn't have proper lock coverage today */
1651         BUG_ON(ret == -EIOCBQUEUED && !(file->f_flags & O_DIRECT));
1652
1653         /*
1654          * deep in g_f_a_w_n()->ocfs2_direct_IO we pass in a ocfs2_dio_end_io
1655          * function pointer which is called when o_direct io completes so that
1656          * it can unlock our rw lock.  (it's the clustered equivalent of
1657          * i_alloc_sem; protects truncate from racing with pending ios).
1658          * Unfortunately there are error cases which call end_io and others
1659          * that don't.  so we don't have to unlock the rw_lock if either an
1660          * async dio is going to do it in the future or an end_io after an
1661          * error has already done it.
1662          */
1663         if (ret == -EIOCBQUEUED || !ocfs2_iocb_is_rw_locked(iocb)) {
1664                 rw_level = -1;
1665                 have_alloc_sem = 0;
1666         }
1667
1668 out:
1669         if (rw_level != -1)
1670                 ocfs2_rw_unlock(inode, rw_level);
1671
1672 out_sems:
1673         if (have_alloc_sem)
1674                 up_read(&inode->i_alloc_sem);
1675
1676         if (written > 0 && sync) {
1677                 ssize_t err;
1678
1679                 err = sync_page_range_nolock(inode, file->f_mapping, pos, count);
1680                 if (err < 0)
1681                         written = err;
1682         }
1683
1684         mutex_unlock(&inode->i_mutex);
1685
1686         mlog_exit(ret);
1687         return written ? written : ret;
1688 }
1689
1690 static int ocfs2_splice_write_actor(struct pipe_inode_info *pipe,
1691                                     struct pipe_buffer *buf,
1692                                     struct splice_desc *sd)
1693 {
1694         int ret, count;
1695         ssize_t copied = 0;
1696         struct file *file = sd->u.file;
1697         unsigned int offset;
1698         struct page *page = NULL;
1699         void *fsdata;
1700         char *src, *dst;
1701
1702         ret = buf->ops->confirm(pipe, buf);
1703         if (ret)
1704                 goto out;
1705
1706         offset = sd->pos & ~PAGE_CACHE_MASK;
1707         count = sd->len;
1708         if (count + offset > PAGE_CACHE_SIZE)
1709                 count = PAGE_CACHE_SIZE - offset;
1710
1711         ret = ocfs2_write_begin(file, file->f_mapping, sd->pos, count, 0,
1712                                 &page, &fsdata);
1713         if (ret) {
1714                 mlog_errno(ret);
1715                 goto out;
1716         }
1717
1718         src = buf->ops->map(pipe, buf, 1);
1719         dst = kmap_atomic(page, KM_USER1);
1720         memcpy(dst + offset, src + buf->offset, count);
1721         kunmap_atomic(page, KM_USER1);
1722         buf->ops->unmap(pipe, buf, src);
1723
1724         copied = ocfs2_write_end(file, file->f_mapping, sd->pos, count, count,
1725                                  page, fsdata);
1726         if (copied < 0) {
1727                 mlog_errno(copied);
1728                 ret = copied;
1729                 goto out;
1730         }
1731 out:
1732
1733         return copied ? copied : ret;
1734 }
1735
1736 static ssize_t __ocfs2_file_splice_write(struct pipe_inode_info *pipe,
1737                                          struct file *out,
1738                                          loff_t *ppos,
1739                                          size_t len,
1740                                          unsigned int flags)
1741 {
1742         int ret, err;
1743         struct address_space *mapping = out->f_mapping;
1744         struct inode *inode = mapping->host;
1745         struct splice_desc sd = {
1746                 .total_len = len,
1747                 .flags = flags,
1748                 .pos = *ppos,
1749                 .u.file = out,
1750         };
1751
1752         ret = __splice_from_pipe(pipe, &sd, ocfs2_splice_write_actor);
1753         if (ret > 0) {
1754                 *ppos += ret;
1755
1756                 if (unlikely((out->f_flags & O_SYNC) || IS_SYNC(inode))) {
1757                         err = generic_osync_inode(inode, mapping,
1758                                                   OSYNC_METADATA|OSYNC_DATA);
1759                         if (err)
1760                                 ret = err;
1761                 }
1762         }
1763
1764         return ret;
1765 }
1766
1767 static ssize_t ocfs2_file_splice_write(struct pipe_inode_info *pipe,
1768                                        struct file *out,
1769                                        loff_t *ppos,
1770                                        size_t len,
1771                                        unsigned int flags)
1772 {
1773         int ret;
1774         struct inode *inode = out->f_path.dentry->d_inode;
1775
1776         mlog_entry("(0x%p, 0x%p, %u, '%.*s')\n", out, pipe,
1777                    (unsigned int)len,
1778                    out->f_path.dentry->d_name.len,
1779                    out->f_path.dentry->d_name.name);
1780
1781         inode_double_lock(inode, pipe->inode);
1782
1783         ret = ocfs2_rw_lock(inode, 1);
1784         if (ret < 0) {
1785                 mlog_errno(ret);
1786                 goto out;
1787         }
1788
1789         ret = ocfs2_prepare_inode_for_write(out->f_path.dentry, ppos, len, 0,
1790                                             NULL);
1791         if (ret < 0) {
1792                 mlog_errno(ret);
1793                 goto out_unlock;
1794         }
1795
1796         /* ok, we're done with i_size and alloc work */
1797         ret = __ocfs2_file_splice_write(pipe, out, ppos, len, flags);
1798
1799 out_unlock:
1800         ocfs2_rw_unlock(inode, 1);
1801 out:
1802         inode_double_unlock(inode, pipe->inode);
1803
1804         mlog_exit(ret);
1805         return ret;
1806 }
1807
1808 static ssize_t ocfs2_file_splice_read(struct file *in,
1809                                       loff_t *ppos,
1810                                       struct pipe_inode_info *pipe,
1811                                       size_t len,
1812                                       unsigned int flags)
1813 {
1814         int ret = 0;
1815         struct inode *inode = in->f_path.dentry->d_inode;
1816
1817         mlog_entry("(0x%p, 0x%p, %u, '%.*s')\n", in, pipe,
1818                    (unsigned int)len,
1819                    in->f_path.dentry->d_name.len,
1820                    in->f_path.dentry->d_name.name);
1821
1822         /*
1823          * See the comment in ocfs2_file_aio_read()
1824          */
1825         ret = ocfs2_meta_lock(inode, NULL, 0);
1826         if (ret < 0) {
1827                 mlog_errno(ret);
1828                 goto bail;
1829         }
1830         ocfs2_meta_unlock(inode, 0);
1831
1832         ret = generic_file_splice_read(in, ppos, pipe, len, flags);
1833
1834 bail:
1835         mlog_exit(ret);
1836         return ret;
1837 }
1838
1839 static ssize_t ocfs2_file_aio_read(struct kiocb *iocb,
1840                                    const struct iovec *iov,
1841                                    unsigned long nr_segs,
1842                                    loff_t pos)
1843 {
1844         int ret = 0, rw_level = -1, have_alloc_sem = 0, lock_level = 0;
1845         struct file *filp = iocb->ki_filp;
1846         struct inode *inode = filp->f_path.dentry->d_inode;
1847
1848         mlog_entry("(0x%p, %u, '%.*s')\n", filp,
1849                    (unsigned int)nr_segs,
1850                    filp->f_path.dentry->d_name.len,
1851                    filp->f_path.dentry->d_name.name);
1852
1853         if (!inode) {
1854                 ret = -EINVAL;
1855                 mlog_errno(ret);
1856                 goto bail;
1857         }
1858
1859         /*
1860          * buffered reads protect themselves in ->readpage().  O_DIRECT reads
1861          * need locks to protect pending reads from racing with truncate.
1862          */
1863         if (filp->f_flags & O_DIRECT) {
1864                 down_read(&inode->i_alloc_sem);
1865                 have_alloc_sem = 1;
1866
1867                 ret = ocfs2_rw_lock(inode, 0);
1868                 if (ret < 0) {
1869                         mlog_errno(ret);
1870                         goto bail;
1871                 }
1872                 rw_level = 0;
1873                 /* communicate with ocfs2_dio_end_io */
1874                 ocfs2_iocb_set_rw_locked(iocb, rw_level);
1875         }
1876
1877         /*
1878          * We're fine letting folks race truncates and extending
1879          * writes with read across the cluster, just like they can
1880          * locally. Hence no rw_lock during read.
1881          *
1882          * Take and drop the meta data lock to update inode fields
1883          * like i_size. This allows the checks down below
1884          * generic_file_aio_read() a chance of actually working.
1885          */
1886         ret = ocfs2_meta_lock_atime(inode, filp->f_vfsmnt, &lock_level);
1887         if (ret < 0) {
1888                 mlog_errno(ret);
1889                 goto bail;
1890         }
1891         ocfs2_meta_unlock(inode, lock_level);
1892
1893         ret = generic_file_aio_read(iocb, iov, nr_segs, iocb->ki_pos);
1894         if (ret == -EINVAL)
1895                 mlog(ML_ERROR, "generic_file_aio_read returned -EINVAL\n");
1896
1897         /* buffered aio wouldn't have proper lock coverage today */
1898         BUG_ON(ret == -EIOCBQUEUED && !(filp->f_flags & O_DIRECT));
1899
1900         /* see ocfs2_file_aio_write */
1901         if (ret == -EIOCBQUEUED || !ocfs2_iocb_is_rw_locked(iocb)) {
1902                 rw_level = -1;
1903                 have_alloc_sem = 0;
1904         }
1905
1906 bail:
1907         if (have_alloc_sem)
1908                 up_read(&inode->i_alloc_sem);
1909         if (rw_level != -1)
1910                 ocfs2_rw_unlock(inode, rw_level);
1911         mlog_exit(ret);
1912
1913         return ret;
1914 }
1915
1916 const struct inode_operations ocfs2_file_iops = {
1917         .setattr        = ocfs2_setattr,
1918         .getattr        = ocfs2_getattr,
1919         .permission     = ocfs2_permission,
1920 };
1921
1922 const struct inode_operations ocfs2_special_file_iops = {
1923         .setattr        = ocfs2_setattr,
1924         .getattr        = ocfs2_getattr,
1925         .permission     = ocfs2_permission,
1926 };
1927
1928 const struct file_operations ocfs2_fops = {
1929         .read           = do_sync_read,
1930         .write          = do_sync_write,
1931         .mmap           = ocfs2_mmap,
1932         .fsync          = ocfs2_sync_file,
1933         .release        = ocfs2_file_release,
1934         .open           = ocfs2_file_open,
1935         .aio_read       = ocfs2_file_aio_read,
1936         .aio_write      = ocfs2_file_aio_write,
1937         .ioctl          = ocfs2_ioctl,
1938 #ifdef CONFIG_COMPAT
1939         .compat_ioctl   = ocfs2_compat_ioctl,
1940 #endif
1941         .splice_read    = ocfs2_file_splice_read,
1942         .splice_write   = ocfs2_file_splice_write,
1943 };
1944
1945 const struct file_operations ocfs2_dops = {
1946         .read           = generic_read_dir,
1947         .readdir        = ocfs2_readdir,
1948         .fsync          = ocfs2_sync_file,
1949         .ioctl          = ocfs2_ioctl,
1950 #ifdef CONFIG_COMPAT
1951         .compat_ioctl   = ocfs2_compat_ioctl,
1952 #endif
1953 };