fs/ocfs2/file.c

   1 /* -*- mode: c; c-basic-offset: 8; -*-
   2  * vim: noexpandtab sw=8 ts=8 sts=0:
   3  *
   4  * file.c
   5  *
   6  * File open, close, extend, truncate
   7  *
   8  * Copyright (C) 2002, 2004 Oracle.  All rights reserved.
   9  *
  10  * This program is free software; you can redistribute it and/or
  11  * modify it under the terms of the GNU General Public
  12  * License as published by the Free Software Foundation; either
  13  * version 2 of the License, or (at your option) any later version.
  14  *
  15  * This program is distributed in the hope that it will be useful,
  16  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  17  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  18  * General Public License for more details.
  19  *
  20  * You should have received a copy of the GNU General Public
  21  * License along with this program; if not, write to the
  22  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  23  * Boston, MA 021110-1307, USA.
  24  */
  25
  26 #include <linux/capability.h>
  27 #include <linux/fs.h>
  28 #include <linux/types.h>
  29 #include <linux/slab.h>
  30 #include <linux/highmem.h>
  31 #include <linux/pagemap.h>
  32 #include <linux/uio.h>
  33 #include <linux/sched.h>
  34 #include <linux/splice.h>
  35 #include <linux/mount.h>
  36 #include <linux/writeback.h>
  37 #include <linux/falloc.h>
  38 #include <linux/quotaops.h>
  39 #include <linux/blkdev.h>
  40 #include <linux/backing-dev.h>
  41
  42 #include <cluster/masklog.h>
  43
  44 #include "ocfs2.h"
  45
  46 #include "alloc.h"
  47 #include "aops.h"
  48 #include "dir.h"
  49 #include "dlmglue.h"
  50 #include "extent_map.h"
  51 #include "file.h"
  52 #include "sysfile.h"
  53 #include "inode.h"
  54 #include "ioctl.h"
  55 #include "journal.h"
  56 #include "locks.h"
  57 #include "mmap.h"
  58 #include "suballoc.h"
  59 #include "super.h"
  60 #include "xattr.h"
  61 #include "acl.h"
  62 #include "quota.h"
  63 #include "refcounttree.h"
  64 #include "ocfs2_trace.h"
  65
  66 #include "buffer_head_io.h"
  67
  68 static int ocfs2_init_file_private(struct inode *inode, struct file *file)
  69 {
  70         struct ocfs2_file_private *fp;
  71
  72         fp = kzalloc(sizeof(struct ocfs2_file_private), GFP_KERNEL);
  73         if (!fp)
  74                 return -ENOMEM;
  75
  76         fp->fp_file = file;
  77         mutex_init(&fp->fp_mutex);
  78         ocfs2_file_lock_res_init(&fp->fp_flock, fp);
  79         file->private_data = fp;
  80
  81         return 0;
  82 }
  83
  84 static void ocfs2_free_file_private(struct inode *inode, struct file *file)
  85 {
  86         struct ocfs2_file_private *fp = file->private_data;
  87         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
  88
  89         if (fp) {
  90                 ocfs2_simple_drop_lockres(osb, &fp->fp_flock);
  91                 ocfs2_lock_res_free(&fp->fp_flock);
  92                 kfree(fp);
  93                 file->private_data = NULL;
  94         }
  95 }
  96
  97 static int ocfs2_file_open(struct inode *inode, struct file *file)
  98 {
  99         int status;
 100         int mode = file->f_flags;
 101         struct ocfs2_inode_info *oi = OCFS2_I(inode);
 102
 103         trace_ocfs2_file_open(inode, file, file->f_path.dentry,
 104                               (unsigned long long)OCFS2_I(inode)->ip_blkno,
 105                               file->f_path.dentry->d_name.len,
 106                               file->f_path.dentry->d_name.name, mode);
 107
 108         if (file->f_mode & FMODE_WRITE) {
 109                 status = dquot_initialize(inode);
 110                 if (status)
 111                         goto leave;
 112         }
 113
 114         spin_lock(&oi->ip_lock);
 115
 116         /* Check that the inode hasn't been wiped from disk by another
 117          * node. If it hasn't then we're safe as long as we hold the
 118          * spin lock until our increment of open count. */
 119         if (OCFS2_I(inode)->ip_flags & OCFS2_INODE_DELETED) {
 120                 spin_unlock(&oi->ip_lock);
 121
 122                 status = -ENOENT;
 123                 goto leave;
 124         }
 125
 126         if (mode & O_DIRECT)
 127                 oi->ip_flags |= OCFS2_INODE_OPEN_DIRECT;
 128
 129         oi->ip_open_count++;
 130         spin_unlock(&oi->ip_lock);
 131
 132         status = ocfs2_init_file_private(inode, file);
 133         if (status) {
 134                 /*
 135                  * We want to set open count back if we're failing the
 136                  * open.
 137                  */
 138                 spin_lock(&oi->ip_lock);
 139                 oi->ip_open_count--;
 140                 spin_unlock(&oi->ip_lock);
 141         }
 142
 143 leave:
 144         return status;
 145 }
 146
 147 static int ocfs2_file_release(struct inode *inode, struct file *file)
 148 {
 149         struct ocfs2_inode_info *oi = OCFS2_I(inode);
 150
 151         spin_lock(&oi->ip_lock);
 152         if (!--oi->ip_open_count)
 153                 oi->ip_flags &= ~OCFS2_INODE_OPEN_DIRECT;
 154
 155         trace_ocfs2_file_release(inode, file, file->f_path.dentry,
 156                                  oi->ip_blkno,
 157                                  file->f_path.dentry->d_name.len,
 158                                  file->f_path.dentry->d_name.name,
 159                                  oi->ip_open_count);
 160         spin_unlock(&oi->ip_lock);
 161
 162         ocfs2_free_file_private(inode, file);
 163
 164         return 0;
 165 }
 166
 167 static int ocfs2_dir_open(struct inode *inode, struct file *file)
 168 {
 169         return ocfs2_init_file_private(inode, file);
 170 }
 171
 172 static int ocfs2_dir_release(struct inode *inode, struct file *file)
 173 {
 174         ocfs2_free_file_private(inode, file);
 175         return 0;
 176 }
 177
 178 static int ocfs2_sync_file(struct file *file, loff_t start, loff_t end,
 179                            int datasync)
 180 {
 181         int err = 0;
 182         struct inode *inode = file->f_mapping->host;
 183         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 184         struct ocfs2_inode_info *oi = OCFS2_I(inode);
 185         journal_t *journal = osb->journal->j_journal;
 186         int ret;
 187         tid_t commit_tid;
 188         bool needs_barrier = false;
 189
 190         trace_ocfs2_sync_file(inode, file, file->f_path.dentry,
 191                               OCFS2_I(inode)->ip_blkno,
 192                               file->f_path.dentry->d_name.len,
 193                               file->f_path.dentry->d_name.name,
 194                               (unsigned long long)datasync);
 195
 196         if (ocfs2_is_hard_readonly(osb) || ocfs2_is_soft_readonly(osb))
 197                 return -EROFS;
 198
 199         err = filemap_write_and_wait_range(inode->i_mapping, start, end);
 200         if (err)
 201                 return err;
 202
 203         commit_tid = datasync ? oi->i_datasync_tid : oi->i_sync_tid;
 204         if (journal->j_flags & JBD2_BARRIER &&
 205             !jbd2_trans_will_send_data_barrier(journal, commit_tid))
 206                 needs_barrier = true;
 207         err = jbd2_complete_transaction(journal, commit_tid);
 208         if (needs_barrier) {
 209                 ret = blkdev_issue_flush(inode->i_sb->s_bdev, GFP_KERNEL, NULL);
 210                 if (!err)
 211                         err = ret;
 212         }
 213
 214         if (err)
 215                 mlog_errno(err);
 216
 217         return (err < 0) ? -EIO : 0;
 218 }
 219
 220 int ocfs2_should_update_atime(struct inode *inode,
 221                               struct vfsmount *vfsmnt)
 222 {
 223         struct timespec now;
 224         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 225
 226         if (ocfs2_is_hard_readonly(osb) || ocfs2_is_soft_readonly(osb))
 227                 return 0;
 228
 229         if ((inode->i_flags & S_NOATIME) ||
 230             ((inode->i_sb->s_flags & MS_NODIRATIME) && S_ISDIR(inode->i_mode)))
 231                 return 0;
 232
 233         /*
 234          * We can be called with no vfsmnt structure - NFSD will
 235          * sometimes do this.
 236          *
 237          * Note that our action here is different than touch_atime() -
 238          * if we can't tell whether this is a noatime mount, then we
 239          * don't know whether to trust the value of s_atime_quantum.
 240          */
 241         if (vfsmnt == NULL)
 242                 return 0;
 243
 244         if ((vfsmnt->mnt_flags & MNT_NOATIME) ||
 245             ((vfsmnt->mnt_flags & MNT_NODIRATIME) && S_ISDIR(inode->i_mode)))
 246                 return 0;
 247
 248         if (vfsmnt->mnt_flags & MNT_RELATIME) {
 249                 if ((timespec_compare(&inode->i_atime, &inode->i_mtime) <= 0) ||
 250                     (timespec_compare(&inode->i_atime, &inode->i_ctime) <= 0))
 251                         return 1;
 252
 253                 return 0;
 254         }
 255
 256         now = current_time(inode);
 257         if ((now.tv_sec - inode->i_atime.tv_sec <= osb->s_atime_quantum))
 258                 return 0;
 259         else
 260                 return 1;
 261 }
 262
 263 int ocfs2_update_inode_atime(struct inode *inode,
 264                              struct buffer_head *bh)
 265 {
 266         int ret;
 267         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 268         handle_t *handle;
 269         struct ocfs2_dinode *di = (struct ocfs2_dinode *) bh->b_data;
 270
 271         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
 272         if (IS_ERR(handle)) {
 273                 ret = PTR_ERR(handle);
 274                 mlog_errno(ret);
 275                 goto out;
 276         }
 277
 278         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), bh,
 279                                       OCFS2_JOURNAL_ACCESS_WRITE);
 280         if (ret) {
 281                 mlog_errno(ret);
 282                 goto out_commit;
 283         }
 284
 285         /*
 286          * Don't use ocfs2_mark_inode_dirty() here as we don't always
 287          * have i_mutex to guard against concurrent changes to other
 288          * inode fields.
 289          */
 290         inode->i_atime = current_time(inode);
 291         di->i_atime = cpu_to_le64(inode->i_atime.tv_sec);
 292         di->i_atime_nsec = cpu_to_le32(inode->i_atime.tv_nsec);
 293         ocfs2_update_inode_fsync_trans(handle, inode, 0);
 294         ocfs2_journal_dirty(handle, bh);
 295
 296 out_commit:
 297         ocfs2_commit_trans(OCFS2_SB(inode->i_sb), handle);
 298 out:
 299         return ret;
 300 }
 301
 302 int ocfs2_set_inode_size(handle_t *handle,
 303                                 struct inode *inode,
 304                                 struct buffer_head *fe_bh,
 305                                 u64 new_i_size)
 306 {
 307         int status;
 308
 309         i_size_write(inode, new_i_size);
 310         inode->i_blocks = ocfs2_inode_sector_count(inode);
 311         inode->i_ctime = inode->i_mtime = current_time(inode);
 312
 313         status = ocfs2_mark_inode_dirty(handle, inode, fe_bh);
 314         if (status < 0) {
 315                 mlog_errno(status);
 316                 goto bail;
 317         }
 318
 319 bail:
 320         return status;
 321 }
 322
 323 int ocfs2_simple_size_update(struct inode *inode,
 324                              struct buffer_head *di_bh,
 325                              u64 new_i_size)
 326 {
 327         int ret;
 328         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 329         handle_t *handle = NULL;
 330
 331         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
 332         if (IS_ERR(handle)) {
 333                 ret = PTR_ERR(handle);
 334                 mlog_errno(ret);
 335                 goto out;
 336         }
 337
 338         ret = ocfs2_set_inode_size(handle, inode, di_bh,
 339                                    new_i_size);
 340         if (ret < 0)
 341                 mlog_errno(ret);
 342
 343         ocfs2_update_inode_fsync_trans(handle, inode, 0);
 344         ocfs2_commit_trans(osb, handle);
 345 out:
 346         return ret;
 347 }
 348
 349 static int ocfs2_cow_file_pos(struct inode *inode,
 350                               struct buffer_head *fe_bh,
 351                               u64 offset)
 352 {
 353         int status;
 354         u32 phys, cpos = offset >> OCFS2_SB(inode->i_sb)->s_clustersize_bits;
 355         unsigned int num_clusters = 0;
 356         unsigned int ext_flags = 0;
 357
 358         /*
 359          * If the new offset is aligned to the range of the cluster, there is
 360          * no space for ocfs2_zero_range_for_truncate to fill, so no need to
 361          * CoW either.
 362          */
 363         if ((offset & (OCFS2_SB(inode->i_sb)->s_clustersize - 1)) == 0)
 364                 return 0;
 365
 366         status = ocfs2_get_clusters(inode, cpos, &phys,
 367                                     &num_clusters, &ext_flags);
 368         if (status) {
 369                 mlog_errno(status);
 370                 goto out;
 371         }
 372
 373         if (!(ext_flags & OCFS2_EXT_REFCOUNTED))
 374                 goto out;
 375
 376         return ocfs2_refcount_cow(inode, fe_bh, cpos, 1, cpos+1);
 377
 378 out:
 379         return status;
 380 }
 381
 382 static int ocfs2_orphan_for_truncate(struct ocfs2_super *osb,
 383                                      struct inode *inode,
 384                                      struct buffer_head *fe_bh,
 385                                      u64 new_i_size)
 386 {
 387         int status;
 388         handle_t *handle;
 389         struct ocfs2_dinode *di;
 390         u64 cluster_bytes;
 391
 392         /*
 393          * We need to CoW the cluster contains the offset if it is reflinked
 394          * since we will call ocfs2_zero_range_for_truncate later which will
 395          * write "0" from offset to the end of the cluster.
 396          */
 397         status = ocfs2_cow_file_pos(inode, fe_bh, new_i_size);
 398         if (status) {
 399                 mlog_errno(status);
 400                 return status;
 401         }
 402
 403         /* TODO: This needs to actually orphan the inode in this
 404          * transaction. */
 405
 406         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
 407         if (IS_ERR(handle)) {
 408                 status = PTR_ERR(handle);
 409                 mlog_errno(status);
 410                 goto out;
 411         }
 412
 413         status = ocfs2_journal_access_di(handle, INODE_CACHE(inode), fe_bh,
 414                                          OCFS2_JOURNAL_ACCESS_WRITE);
 415         if (status < 0) {
 416                 mlog_errno(status);
 417                 goto out_commit;
 418         }
 419
 420         /*
 421          * Do this before setting i_size.
 422          */
 423         cluster_bytes = ocfs2_align_bytes_to_clusters(inode->i_sb, new_i_size);
 424         status = ocfs2_zero_range_for_truncate(inode, handle, new_i_size,
 425                                                cluster_bytes);
 426         if (status) {
 427                 mlog_errno(status);
 428                 goto out_commit;
 429         }
 430
 431         i_size_write(inode, new_i_size);
 432         inode->i_ctime = inode->i_mtime = current_time(inode);
 433
 434         di = (struct ocfs2_dinode *) fe_bh->b_data;
 435         di->i_size = cpu_to_le64(new_i_size);
 436         di->i_ctime = di->i_mtime = cpu_to_le64(inode->i_ctime.tv_sec);
 437         di->i_ctime_nsec = di->i_mtime_nsec = cpu_to_le32(inode->i_ctime.tv_nsec);
 438         ocfs2_update_inode_fsync_trans(handle, inode, 0);
 439
 440         ocfs2_journal_dirty(handle, fe_bh);
 441
 442 out_commit:
 443         ocfs2_commit_trans(osb, handle);
 444 out:
 445         return status;
 446 }
 447
 448 int ocfs2_truncate_file(struct inode *inode,
 449                                struct buffer_head *di_bh,
 450                                u64 new_i_size)
 451 {
 452         int status = 0;
 453         struct ocfs2_dinode *fe = NULL;
 454         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 455
 456         /* We trust di_bh because it comes from ocfs2_inode_lock(), which
 457          * already validated it */
 458         fe = (struct ocfs2_dinode *) di_bh->b_data;
 459
 460         trace_ocfs2_truncate_file((unsigned long long)OCFS2_I(inode)->ip_blkno,
 461                                   (unsigned long long)le64_to_cpu(fe->i_size),
 462                                   (unsigned long long)new_i_size);
 463
 464         mlog_bug_on_msg(le64_to_cpu(fe->i_size) != i_size_read(inode),
 465                         "Inode %llu, inode i_size = %lld != di "
 466                         "i_size = %llu, i_flags = 0x%x\n",
 467                         (unsigned long long)OCFS2_I(inode)->ip_blkno,
 468                         i_size_read(inode),
 469                         (unsigned long long)le64_to_cpu(fe->i_size),
 470                         le32_to_cpu(fe->i_flags));
 471
 472         if (new_i_size > le64_to_cpu(fe->i_size)) {
 473                 trace_ocfs2_truncate_file_error(
 474                         (unsigned long long)le64_to_cpu(fe->i_size),
 475                         (unsigned long long)new_i_size);
 476                 status = -EINVAL;
 477                 mlog_errno(status);
 478                 goto bail;
 479         }
 480
 481         down_write(&OCFS2_I(inode)->ip_alloc_sem);
 482
 483         ocfs2_resv_discard(&osb->osb_la_resmap,
 484                            &OCFS2_I(inode)->ip_la_data_resv);
 485
 486         /*
 487          * The inode lock forced other nodes to sync and drop their
 488          * pages, which (correctly) happens even if we have a truncate
 489          * without allocation change - ocfs2 cluster sizes can be much
 490          * greater than page size, so we have to truncate them
 491          * anyway.
 492          */
 493         unmap_mapping_range(inode->i_mapping, new_i_size + PAGE_SIZE - 1, 0, 1);
 494         truncate_inode_pages(inode->i_mapping, new_i_size);
 495
 496         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
 497                 status = ocfs2_truncate_inline(inode, di_bh, new_i_size,
 498                                                i_size_read(inode), 1);
 499                 if (status)
 500                         mlog_errno(status);
 501
 502                 goto bail_unlock_sem;
 503         }
 504
 505         /* alright, we're going to need to do a full blown alloc size
 506          * change. Orphan the inode so that recovery can complete the
 507          * truncate if necessary. This does the task of marking
 508          * i_size. */
 509         status = ocfs2_orphan_for_truncate(osb, inode, di_bh, new_i_size);
 510         if (status < 0) {
 511                 mlog_errno(status);
 512                 goto bail_unlock_sem;
 513         }
 514
 515         status = ocfs2_commit_truncate(osb, inode, di_bh);
 516         if (status < 0) {
 517                 mlog_errno(status);
 518                 goto bail_unlock_sem;
 519         }
 520
 521         /* TODO: orphan dir cleanup here. */
 522 bail_unlock_sem:
 523         up_write(&OCFS2_I(inode)->ip_alloc_sem);
 524
 525 bail:
 526         if (!status && OCFS2_I(inode)->ip_clusters == 0)
 527                 status = ocfs2_try_remove_refcount_tree(inode, di_bh);
 528
 529         return status;
 530 }
 531
 532 /*
 533  * extend file allocation only here.
 534  * we'll update all the disk stuff, and oip->alloc_size
 535  *
 536  * expect stuff to be locked, a transaction started and enough data /
 537  * metadata reservations in the contexts.
 538  *
 539  * Will return -EAGAIN, and a reason if a restart is needed.
 540  * If passed in, *reason will always be set, even in error.
 541  */
 542 int ocfs2_add_inode_data(struct ocfs2_super *osb,
 543                          struct inode *inode,
 544                          u32 *logical_offset,
 545                          u32 clusters_to_add,
 546                          int mark_unwritten,
 547                          struct buffer_head *fe_bh,
 548                          handle_t *handle,
 549                          struct ocfs2_alloc_context *data_ac,
 550                          struct ocfs2_alloc_context *meta_ac,
 551                          enum ocfs2_alloc_restarted *reason_ret)
 552 {
 553         int ret;
 554         struct ocfs2_extent_tree et;
 555
 556         ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), fe_bh);
 557         ret = ocfs2_add_clusters_in_btree(handle, &et, logical_offset,
 558                                           clusters_to_add, mark_unwritten,
 559                                           data_ac, meta_ac, reason_ret);
 560
 561         return ret;
 562 }
 563
 564 static int __ocfs2_extend_allocation(struct inode *inode, u32 logical_start,
 565                                      u32 clusters_to_add, int mark_unwritten)
 566 {
 567         int status = 0;
 568         int restart_func = 0;
 569         int credits;
 570         u32 prev_clusters;
 571         struct buffer_head *bh = NULL;
 572         struct ocfs2_dinode *fe = NULL;
 573         handle_t *handle = NULL;
 574         struct ocfs2_alloc_context *data_ac = NULL;
 575         struct ocfs2_alloc_context *meta_ac = NULL;
 576         enum ocfs2_alloc_restarted why = RESTART_NONE;
 577         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 578         struct ocfs2_extent_tree et;
 579         int did_quota = 0;
 580
 581         /*
 582          * Unwritten extent only exists for file systems which
 583          * support holes.
 584          */
 585         BUG_ON(mark_unwritten && !ocfs2_sparse_alloc(osb));
 586
 587         status = ocfs2_read_inode_block(inode, &bh);
 588         if (status < 0) {
 589                 mlog_errno(status);
 590                 goto leave;
 591         }
 592         fe = (struct ocfs2_dinode *) bh->b_data;
 593
 594 restart_all:
 595         BUG_ON(le32_to_cpu(fe->i_clusters) != OCFS2_I(inode)->ip_clusters);
 596
 597         ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), bh);
 598         status = ocfs2_lock_allocators(inode, &et, clusters_to_add, 0,
 599                                        &data_ac, &meta_ac);
 600         if (status) {
 601                 mlog_errno(status);
 602                 goto leave;
 603         }
 604
 605         credits = ocfs2_calc_extend_credits(osb->sb, &fe->id2.i_list);
 606         handle = ocfs2_start_trans(osb, credits);
 607         if (IS_ERR(handle)) {
 608                 status = PTR_ERR(handle);
 609                 handle = NULL;
 610                 mlog_errno(status);
 611                 goto leave;
 612         }
 613
 614 restarted_transaction:
 615         trace_ocfs2_extend_allocation(
 616                 (unsigned long long)OCFS2_I(inode)->ip_blkno,
 617                 (unsigned long long)i_size_read(inode),
 618                 le32_to_cpu(fe->i_clusters), clusters_to_add,
 619                 why, restart_func);
 620
 621         status = dquot_alloc_space_nodirty(inode,
 622                         ocfs2_clusters_to_bytes(osb->sb, clusters_to_add));
 623         if (status)
 624                 goto leave;
 625         did_quota = 1;
 626
 627         /* reserve a write to the file entry early on - that we if we
 628          * run out of credits in the allocation path, we can still
 629          * update i_size. */
 630         status = ocfs2_journal_access_di(handle, INODE_CACHE(inode), bh,
 631                                          OCFS2_JOURNAL_ACCESS_WRITE);
 632         if (status < 0) {
 633                 mlog_errno(status);
 634                 goto leave;
 635         }
 636
 637         prev_clusters = OCFS2_I(inode)->ip_clusters;
 638
 639         status = ocfs2_add_inode_data(osb,
 640                                       inode,
 641                                       &logical_start,
 642                                       clusters_to_add,
 643                                       mark_unwritten,
 644                                       bh,
 645                                       handle,
 646                                       data_ac,
 647                                       meta_ac,
 648                                       &why);
 649         if ((status < 0) && (status != -EAGAIN)) {
 650                 if (status != -ENOSPC)
 651                         mlog_errno(status);
 652                 goto leave;
 653         }
 654         ocfs2_update_inode_fsync_trans(handle, inode, 1);
 655         ocfs2_journal_dirty(handle, bh);
 656
 657         spin_lock(&OCFS2_I(inode)->ip_lock);
 658         clusters_to_add -= (OCFS2_I(inode)->ip_clusters - prev_clusters);
 659         spin_unlock(&OCFS2_I(inode)->ip_lock);
 660         /* Release unused quota reservation */
 661         dquot_free_space(inode,
 662                         ocfs2_clusters_to_bytes(osb->sb, clusters_to_add));
 663         did_quota = 0;
 664
 665         if (why != RESTART_NONE && clusters_to_add) {
 666                 if (why == RESTART_META) {
 667                         restart_func = 1;
 668                         status = 0;
 669                 } else {
 670                         BUG_ON(why != RESTART_TRANS);
 671
 672                         status = ocfs2_allocate_extend_trans(handle, 1);
 673                         if (status < 0) {
 674                                 /* handle still has to be committed at
 675                                  * this point. */
 676                                 status = -ENOMEM;
 677                                 mlog_errno(status);
 678                                 goto leave;
 679                         }
 680                         goto restarted_transaction;
 681                 }
 682         }
 683
 684         trace_ocfs2_extend_allocation_end(OCFS2_I(inode)->ip_blkno,
 685              le32_to_cpu(fe->i_clusters),
 686              (unsigned long long)le64_to_cpu(fe->i_size),
 687              OCFS2_I(inode)->ip_clusters,
 688              (unsigned long long)i_size_read(inode));
 689
 690 leave:
 691         if (status < 0 && did_quota)
 692                 dquot_free_space(inode,
 693                         ocfs2_clusters_to_bytes(osb->sb, clusters_to_add));
 694         if (handle) {
 695                 ocfs2_commit_trans(osb, handle);
 696                 handle = NULL;
 697         }
 698         if (data_ac) {
 699                 ocfs2_free_alloc_context(data_ac);
 700                 data_ac = NULL;
 701         }
 702         if (meta_ac) {
 703                 ocfs2_free_alloc_context(meta_ac);
 704                 meta_ac = NULL;
 705         }
 706         if ((!status) && restart_func) {
 707                 restart_func = 0;
 708                 goto restart_all;
 709         }
 710         brelse(bh);
 711         bh = NULL;
 712
 713         return status;
 714 }
 715
 716 int ocfs2_extend_allocation(struct inode *inode, u32 logical_start,
 717                 u32 clusters_to_add, int mark_unwritten)
 718 {
 719         return __ocfs2_extend_allocation(inode, logical_start,
 720                         clusters_to_add, mark_unwritten);
 721 }
 722
 723 /*
 724  * While a write will already be ordering the data, a truncate will not.
 725  * Thus, we need to explicitly order the zeroed pages.
 726  */
 727 static handle_t *ocfs2_zero_start_ordered_transaction(struct inode *inode,
 728                                                 struct buffer_head *di_bh)
 729 {
 730         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 731         handle_t *handle = NULL;
 732         int ret = 0;
 733
 734         if (!ocfs2_should_order_data(inode))
 735                 goto out;
 736
 737         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
 738         if (IS_ERR(handle)) {
 739                 ret = -ENOMEM;
 740                 mlog_errno(ret);
 741                 goto out;
 742         }
 743
 744         ret = ocfs2_jbd2_file_inode(handle, inode);
 745         if (ret < 0) {
 746                 mlog_errno(ret);
 747                 goto out;
 748         }
 749
 750         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
 751                                       OCFS2_JOURNAL_ACCESS_WRITE);
 752         if (ret)
 753                 mlog_errno(ret);
 754         ocfs2_update_inode_fsync_trans(handle, inode, 1);
 755
 756 out:
 757         if (ret) {
 758                 if (!IS_ERR(handle))
 759                         ocfs2_commit_trans(osb, handle);
 760                 handle = ERR_PTR(ret);
 761         }
 762         return handle;
 763 }
 764
 765 /* Some parts of this taken from generic_cont_expand, which turned out
 766  * to be too fragile to do exactly what we need without us having to
 767  * worry about recursive locking in ->write_begin() and ->write_end(). */
 768 static int ocfs2_write_zero_page(struct inode *inode, u64 abs_from,
 769                                  u64 abs_to, struct buffer_head *di_bh)
 770 {
 771         struct address_space *mapping = inode->i_mapping;
 772         struct page *page;
 773         unsigned long index = abs_from >> PAGE_SHIFT;
 774         handle_t *handle;
 775         int ret = 0;
 776         unsigned zero_from, zero_to, block_start, block_end;
 777         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
 778
 779         BUG_ON(abs_from >= abs_to);
 780         BUG_ON(abs_to > (((u64)index + 1) << PAGE_SHIFT));
 781         BUG_ON(abs_from & (inode->i_blkbits - 1));
 782
 783         handle = ocfs2_zero_start_ordered_transaction(inode, di_bh);
 784         if (IS_ERR(handle)) {
 785                 ret = PTR_ERR(handle);
 786                 goto out;
 787         }
 788
 789         page = find_or_create_page(mapping, index, GFP_NOFS);
 790         if (!page) {
 791                 ret = -ENOMEM;
 792                 mlog_errno(ret);
 793                 goto out_commit_trans;
 794         }
 795
 796         /* Get the offsets within the page that we want to zero */
 797         zero_from = abs_from & (PAGE_SIZE - 1);
 798         zero_to = abs_to & (PAGE_SIZE - 1);
 799         if (!zero_to)
 800                 zero_to = PAGE_SIZE;
 801
 802         trace_ocfs2_write_zero_page(
 803                         (unsigned long long)OCFS2_I(inode)->ip_blkno,
 804                         (unsigned long long)abs_from,
 805                         (unsigned long long)abs_to,
 806                         index, zero_from, zero_to);
 807
 808         /* We know that zero_from is block aligned */
 809         for (block_start = zero_from; block_start < zero_to;
 810              block_start = block_end) {
 811                 block_end = block_start + i_blocksize(inode);
 812
 813                 /*
 814                  * block_start is block-aligned.  Bump it by one to force
 815                  * __block_write_begin and block_commit_write to zero the
 816                  * whole block.
 817                  */
 818                 ret = __block_write_begin(page, block_start + 1, 0,
 819                                           ocfs2_get_block);
 820                 if (ret < 0) {
 821                         mlog_errno(ret);
 822                         goto out_unlock;
 823                 }
 824
 825
 826                 /* must not update i_size! */
 827                 ret = block_commit_write(page, block_start + 1,
 828                                          block_start + 1);
 829                 if (ret < 0)
 830                         mlog_errno(ret);
 831                 else
 832                         ret = 0;
 833         }
 834
 835         /*
 836          * fs-writeback will release the dirty pages without page lock
 837          * whose offset are over inode size, the release happens at
 838          * block_write_full_page().
 839          */
 840         i_size_write(inode, abs_to);
 841         inode->i_blocks = ocfs2_inode_sector_count(inode);
 842         di->i_size = cpu_to_le64((u64)i_size_read(inode));
 843         inode->i_mtime = inode->i_ctime = current_time(inode);
 844         di->i_mtime = di->i_ctime = cpu_to_le64(inode->i_mtime.tv_sec);
 845         di->i_ctime_nsec = cpu_to_le32(inode->i_mtime.tv_nsec);
 846         di->i_mtime_nsec = di->i_ctime_nsec;
 847         if (handle) {
 848                 ocfs2_journal_dirty(handle, di_bh);
 849                 ocfs2_update_inode_fsync_trans(handle, inode, 1);
 850         }
 851
 852 out_unlock:
 853         unlock_page(page);
 854         put_page(page);
 855 out_commit_trans:
 856         if (handle)
 857                 ocfs2_commit_trans(OCFS2_SB(inode->i_sb), handle);
 858 out:
 859         return ret;
 860 }
 861
 862 /*
 863  * Find the next range to zero.  We do this in terms of bytes because
 864  * that's what ocfs2_zero_extend() wants, and it is dealing with the
 865  * pagecache.  We may return multiple extents.
 866  *
 867  * zero_start and zero_end are ocfs2_zero_extend()s current idea of what
 868  * needs to be zeroed.  range_start and range_end return the next zeroing
 869  * range.  A subsequent call should pass the previous range_end as its
 870  * zero_start.  If range_end is 0, there's nothing to do.
 871  *
 872  * Unwritten extents are skipped over.  Refcounted extents are CoWd.
 873  */
 874 static int ocfs2_zero_extend_get_range(struct inode *inode,
 875                                        struct buffer_head *di_bh,
 876                                        u64 zero_start, u64 zero_end,
 877                                        u64 *range_start, u64 *range_end)
 878 {
 879         int rc = 0, needs_cow = 0;
 880         u32 p_cpos, zero_clusters = 0;
 881         u32 zero_cpos =
 882                 zero_start >> OCFS2_SB(inode->i_sb)->s_clustersize_bits;
 883         u32 last_cpos = ocfs2_clusters_for_bytes(inode->i_sb, zero_end);
 884         unsigned int num_clusters = 0;
 885         unsigned int ext_flags = 0;
 886
 887         while (zero_cpos < last_cpos) {
 888                 rc = ocfs2_get_clusters(inode, zero_cpos, &p_cpos,
 889                                         &num_clusters, &ext_flags);
 890                 if (rc) {
 891                         mlog_errno(rc);
 892                         goto out;
 893                 }
 894
 895                 if (p_cpos && !(ext_flags & OCFS2_EXT_UNWRITTEN)) {
 896                         zero_clusters = num_clusters;
 897                         if (ext_flags & OCFS2_EXT_REFCOUNTED)
 898                                 needs_cow = 1;
 899                         break;
 900                 }
 901
 902                 zero_cpos += num_clusters;
 903         }
 904         if (!zero_clusters) {
 905                 *range_end = 0;
 906                 goto out;
 907         }
 908
 909         while ((zero_cpos + zero_clusters) < last_cpos) {
 910                 rc = ocfs2_get_clusters(inode, zero_cpos + zero_clusters,
 911                                         &p_cpos, &num_clusters,
 912                                         &ext_flags);
 913                 if (rc) {
 914                         mlog_errno(rc);
 915                         goto out;
 916                 }
 917
 918                 if (!p_cpos || (ext_flags & OCFS2_EXT_UNWRITTEN))
 919                         break;
 920                 if (ext_flags & OCFS2_EXT_REFCOUNTED)
 921                         needs_cow = 1;
 922                 zero_clusters += num_clusters;
 923         }
 924         if ((zero_cpos + zero_clusters) > last_cpos)
 925                 zero_clusters = last_cpos - zero_cpos;
 926
 927         if (needs_cow) {
 928                 rc = ocfs2_refcount_cow(inode, di_bh, zero_cpos,
 929                                         zero_clusters, UINT_MAX);
 930                 if (rc) {
 931                         mlog_errno(rc);
 932                         goto out;
 933                 }
 934         }
 935
 936         *range_start = ocfs2_clusters_to_bytes(inode->i_sb, zero_cpos);
 937         *range_end = ocfs2_clusters_to_bytes(inode->i_sb,
 938                                              zero_cpos + zero_clusters);
 939
 940 out:
 941         return rc;
 942 }
 943
 944 /*
 945  * Zero one range returned from ocfs2_zero_extend_get_range().  The caller
 946  * has made sure that the entire range needs zeroing.
 947  */
 948 static int ocfs2_zero_extend_range(struct inode *inode, u64 range_start,
 949                                    u64 range_end, struct buffer_head *di_bh)
 950 {
 951         int rc = 0;
 952         u64 next_pos;
 953         u64 zero_pos = range_start;
 954
 955         trace_ocfs2_zero_extend_range(
 956                         (unsigned long long)OCFS2_I(inode)->ip_blkno,
 957                         (unsigned long long)range_start,
 958                         (unsigned long long)range_end);
 959         BUG_ON(range_start >= range_end);
 960
 961         while (zero_pos < range_end) {
 962                 next_pos = (zero_pos & PAGE_MASK) + PAGE_SIZE;
 963                 if (next_pos > range_end)
 964                         next_pos = range_end;
 965                 rc = ocfs2_write_zero_page(inode, zero_pos, next_pos, di_bh);
 966                 if (rc < 0) {
 967                         mlog_errno(rc);
 968                         break;
 969                 }
 970                 zero_pos = next_pos;
 971
 972                 /*
 973                  * Very large extends have the potential to lock up
 974                  * the cpu for extended periods of time.
 975                  */
 976                 cond_resched();
 977         }
 978
 979         return rc;
 980 }
 981
 982 int ocfs2_zero_extend(struct inode *inode, struct buffer_head *di_bh,
 983                       loff_t zero_to_size)
 984 {
 985         int ret = 0;
 986         u64 zero_start, range_start = 0, range_end = 0;
 987         struct super_block *sb = inode->i_sb;
 988
 989         zero_start = ocfs2_align_bytes_to_blocks(sb, i_size_read(inode));
 990         trace_ocfs2_zero_extend((unsigned long long)OCFS2_I(inode)->ip_blkno,
 991                                 (unsigned long long)zero_start,
 992                                 (unsigned long long)i_size_read(inode));
 993         while (zero_start < zero_to_size) {
 994                 ret = ocfs2_zero_extend_get_range(inode, di_bh, zero_start,
 995                                                   zero_to_size,
 996                                                   &range_start,
 997                                                   &range_end);
 998                 if (ret) {
 999                         mlog_errno(ret);
1000                         break;
1001                 }
1002                 if (!range_end)
1003                         break;
1004                 /* Trim the ends */
1005                 if (range_start < zero_start)
1006                         range_start = zero_start;
1007                 if (range_end > zero_to_size)
1008                         range_end = zero_to_size;
1009
1010                 ret = ocfs2_zero_extend_range(inode, range_start,
1011                                               range_end, di_bh);
1012                 if (ret) {
1013                         mlog_errno(ret);
1014                         break;
1015                 }
1016                 zero_start = range_end;
1017         }
1018
1019         return ret;
1020 }
1021
1022 int ocfs2_extend_no_holes(struct inode *inode, struct buffer_head *di_bh,
1023                           u64 new_i_size, u64 zero_to)
1024 {
1025         int ret;
1026         u32 clusters_to_add;
1027         struct ocfs2_inode_info *oi = OCFS2_I(inode);
1028
1029         /*
1030          * Only quota files call this without a bh, and they can't be
1031          * refcounted.
1032          */
1033         BUG_ON(!di_bh && ocfs2_is_refcount_inode(inode));
1034         BUG_ON(!di_bh && !(oi->ip_flags & OCFS2_INODE_SYSTEM_FILE));
1035
1036         clusters_to_add = ocfs2_clusters_for_bytes(inode->i_sb, new_i_size);
1037         if (clusters_to_add < oi->ip_clusters)
1038                 clusters_to_add = 0;
1039         else
1040                 clusters_to_add -= oi->ip_clusters;
1041
1042         if (clusters_to_add) {
1043                 ret = __ocfs2_extend_allocation(inode, oi->ip_clusters,
1044                                                 clusters_to_add, 0);
1045                 if (ret) {
1046                         mlog_errno(ret);
1047                         goto out;
1048                 }
1049         }
1050
1051         /*
1052          * Call this even if we don't add any clusters to the tree. We
1053          * still need to zero the area between the old i_size and the
1054          * new i_size.
1055          */
1056         ret = ocfs2_zero_extend(inode, di_bh, zero_to);
1057         if (ret < 0)
1058                 mlog_errno(ret);
1059
1060 out:
1061         return ret;
1062 }
1063
1064 static int ocfs2_extend_file(struct inode *inode,
1065                              struct buffer_head *di_bh,
1066                              u64 new_i_size)
1067 {
1068         int ret = 0;
1069         struct ocfs2_inode_info *oi = OCFS2_I(inode);
1070
1071         BUG_ON(!di_bh);
1072
1073         /* setattr sometimes calls us like this. */
1074         if (new_i_size == 0)
1075                 goto out;
1076
1077         if (i_size_read(inode) == new_i_size)
1078                 goto out;
1079         BUG_ON(new_i_size < i_size_read(inode));
1080
1081         /*
1082          * The alloc sem blocks people in read/write from reading our
1083          * allocation until we're done changing it. We depend on
1084          * i_mutex to block other extend/truncate calls while we're
1085          * here.  We even have to hold it for sparse files because there
1086          * might be some tail zeroing.
1087          */
1088         down_write(&oi->ip_alloc_sem);
1089
1090         if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1091                 /*
1092                  * We can optimize small extends by keeping the inodes
1093                  * inline data.
1094                  */
1095                 if (ocfs2_size_fits_inline_data(di_bh, new_i_size)) {
1096                         up_write(&oi->ip_alloc_sem);
1097                         goto out_update_size;
1098                 }
1099
1100                 ret = ocfs2_convert_inline_data_to_extents(inode, di_bh);
1101                 if (ret) {
1102                         up_write(&oi->ip_alloc_sem);
1103                         mlog_errno(ret);
1104                         goto out;
1105                 }
1106         }
1107
1108         if (ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)))
1109                 ret = ocfs2_zero_extend(inode, di_bh, new_i_size);
1110         else
1111                 ret = ocfs2_extend_no_holes(inode, di_bh, new_i_size,
1112                                             new_i_size);
1113
1114         up_write(&oi->ip_alloc_sem);
1115
1116         if (ret < 0) {
1117                 mlog_errno(ret);
1118                 goto out;
1119         }
1120
1121 out_update_size:
1122         ret = ocfs2_simple_size_update(inode, di_bh, new_i_size);
1123         if (ret < 0)
1124                 mlog_errno(ret);
1125
1126 out:
1127         return ret;
1128 }
1129
1130 int ocfs2_setattr(struct dentry *dentry, struct iattr *attr)
1131 {
1132         int status = 0, size_change;
1133         int inode_locked = 0;
1134         struct inode *inode = d_inode(dentry);
1135         struct super_block *sb = inode->i_sb;
1136         struct ocfs2_super *osb = OCFS2_SB(sb);
1137         struct buffer_head *bh = NULL;
1138         handle_t *handle = NULL;
1139         struct dquot *transfer_to[MAXQUOTAS] = { };
1140         int qtype;
1141         int had_lock;
1142         struct ocfs2_lock_holder oh;
1143
1144         trace_ocfs2_setattr(inode, dentry,
1145                             (unsigned long long)OCFS2_I(inode)->ip_blkno,
1146                             dentry->d_name.len, dentry->d_name.name,
1147                             attr->ia_valid, attr->ia_mode,
1148                             from_kuid(&init_user_ns, attr->ia_uid),
1149                             from_kgid(&init_user_ns, attr->ia_gid));
1150
1151         /* ensuring we don't even attempt to truncate a symlink */
1152         if (S_ISLNK(inode->i_mode))
1153                 attr->ia_valid &= ~ATTR_SIZE;
1154
1155 #define OCFS2_VALID_ATTRS (ATTR_ATIME | ATTR_MTIME | ATTR_CTIME | ATTR_SIZE \
1156                            | ATTR_GID | ATTR_UID | ATTR_MODE)
1157         if (!(attr->ia_valid & OCFS2_VALID_ATTRS))
1158                 return 0;
1159
1160         status = setattr_prepare(dentry, attr);
1161         if (status)
1162                 return status;
1163
1164         if (is_quota_modification(inode, attr)) {
1165                 status = dquot_initialize(inode);
1166                 if (status)
1167                         return status;
1168         }
1169         size_change = S_ISREG(inode->i_mode) && attr->ia_valid & ATTR_SIZE;
1170         if (size_change) {
1171                 status = ocfs2_rw_lock(inode, 1);
1172                 if (status < 0) {
1173                         mlog_errno(status);
1174                         goto bail;
1175                 }
1176         }
1177
1178         had_lock = ocfs2_inode_lock_tracker(inode, &bh, 1, &oh);
1179         if (had_lock < 0) {
1180                 status = had_lock;
1181                 goto bail_unlock_rw;
1182         } else if (had_lock) {
1183                 /*
1184                  * As far as we know, ocfs2_setattr() could only be the first
1185                  * VFS entry point in the call chain of recursive cluster
1186                  * locking issue.
1187                  *
1188                  * For instance:
1189                  * chmod_common()
1190                  *  notify_change()
1191                  *   ocfs2_setattr()
1192                  *    posix_acl_chmod()
1193                  *     ocfs2_iop_get_acl()
1194                  *
1195                  * But, we're not 100% sure if it's always true, because the
1196                  * ordering of the VFS entry points in the call chain is out
1197                  * of our control. So, we'd better dump the stack here to
1198                  * catch the other cases of recursive locking.
1199                  */
1200                 mlog(ML_ERROR, "Another case of recursive locking:\n");
1201                 dump_stack();
1202         }
1203         inode_locked = 1;
1204
1205         if (size_change) {
1206                 status = inode_newsize_ok(inode, attr->ia_size);
1207                 if (status)
1208                         goto bail_unlock;
1209
1210                 inode_dio_wait(inode);
1211
1212                 if (i_size_read(inode) >= attr->ia_size) {
1213                         if (ocfs2_should_order_data(inode)) {
1214                                 status = ocfs2_begin_ordered_truncate(inode,
1215                                                                       attr->ia_size);
1216                                 if (status)
1217                                         goto bail_unlock;
1218                         }
1219                         status = ocfs2_truncate_file(inode, bh, attr->ia_size);
1220                 } else
1221                         status = ocfs2_extend_file(inode, bh, attr->ia_size);
1222                 if (status < 0) {
1223                         if (status != -ENOSPC)
1224                                 mlog_errno(status);
1225                         status = -ENOSPC;
1226                         goto bail_unlock;
1227                 }
1228         }
1229
1230         if ((attr->ia_valid & ATTR_UID && !uid_eq(attr->ia_uid, inode->i_uid)) ||
1231             (attr->ia_valid & ATTR_GID && !gid_eq(attr->ia_gid, inode->i_gid))) {
1232                 /*
1233                  * Gather pointers to quota structures so that allocation /
1234                  * freeing of quota structures happens here and not inside
1235                  * dquot_transfer() where we have problems with lock ordering
1236                  */
1237                 if (attr->ia_valid & ATTR_UID && !uid_eq(attr->ia_uid, inode->i_uid)
1238                     && OCFS2_HAS_RO_COMPAT_FEATURE(sb,
1239                     OCFS2_FEATURE_RO_COMPAT_USRQUOTA)) {
1240                         transfer_to[USRQUOTA] = dqget(sb, make_kqid_uid(attr->ia_uid));
1241                         if (IS_ERR(transfer_to[USRQUOTA])) {
1242                                 status = PTR_ERR(transfer_to[USRQUOTA]);
1243                                 goto bail_unlock;
1244                         }
1245                 }
1246                 if (attr->ia_valid & ATTR_GID && !gid_eq(attr->ia_gid, inode->i_gid)
1247                     && OCFS2_HAS_RO_COMPAT_FEATURE(sb,
1248                     OCFS2_FEATURE_RO_COMPAT_GRPQUOTA)) {
1249                         transfer_to[GRPQUOTA] = dqget(sb, make_kqid_gid(attr->ia_gid));
1250                         if (IS_ERR(transfer_to[GRPQUOTA])) {
1251                                 status = PTR_ERR(transfer_to[GRPQUOTA]);
1252                                 goto bail_unlock;
1253                         }
1254                 }
1255                 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS +
1256                                            2 * ocfs2_quota_trans_credits(sb));
1257                 if (IS_ERR(handle)) {
1258                         status = PTR_ERR(handle);
1259                         mlog_errno(status);
1260                         goto bail_unlock;
1261                 }
1262                 status = __dquot_transfer(inode, transfer_to);
1263                 if (status < 0)
1264                         goto bail_commit;
1265         } else {
1266                 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1267                 if (IS_ERR(handle)) {
1268                         status = PTR_ERR(handle);
1269                         mlog_errno(status);
1270                         goto bail_unlock;
1271                 }
1272         }
1273
1274         setattr_copy(inode, attr);
1275         mark_inode_dirty(inode);
1276
1277         status = ocfs2_mark_inode_dirty(handle, inode, bh);
1278         if (status < 0)
1279                 mlog_errno(status);
1280
1281 bail_commit:
1282         ocfs2_commit_trans(osb, handle);
1283 bail_unlock:
1284         if (status && inode_locked) {
1285                 ocfs2_inode_unlock_tracker(inode, 1, &oh, had_lock);
1286                 inode_locked = 0;
1287         }
1288 bail_unlock_rw:
1289         if (size_change)
1290                 ocfs2_rw_unlock(inode, 1);
1291 bail:
1292
1293         /* Release quota pointers in case we acquired them */
1294         for (qtype = 0; qtype < OCFS2_MAXQUOTAS; qtype++)
1295                 dqput(transfer_to[qtype]);
1296
1297         if (!status && attr->ia_valid & ATTR_MODE) {
1298                 status = ocfs2_acl_chmod(inode, bh);
1299                 if (status < 0)
1300                         mlog_errno(status);
1301         }
1302         if (inode_locked)
1303                 ocfs2_inode_unlock_tracker(inode, 1, &oh, had_lock);
1304
1305         brelse(bh);
1306         return status;
1307 }
1308
1309 int ocfs2_getattr(struct vfsmount *mnt,
1310                   struct dentry *dentry,
1311                   struct kstat *stat)
1312 {
1313         struct inode *inode = d_inode(dentry);
1314         struct super_block *sb = dentry->d_sb;
1315         struct ocfs2_super *osb = sb->s_fs_info;
1316         int err;
1317
1318         err = ocfs2_inode_revalidate(dentry);
1319         if (err) {
1320                 if (err != -ENOENT)
1321                         mlog_errno(err);
1322                 goto bail;
1323         }
1324
1325         generic_fillattr(inode, stat);
1326         /*
1327          * If there is inline data in the inode, the inode will normally not
1328          * have data blocks allocated (it may have an external xattr block).
1329          * Report at least one sector for such files, so tools like tar, rsync,
1330          * others don't incorrectly think the file is completely sparse.
1331          */
1332         if (unlikely(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL))
1333                 stat->blocks += (stat->size + 511)>>9;
1334
1335         /* We set the blksize from the cluster size for performance */
1336         stat->blksize = osb->s_clustersize;
1337
1338 bail:
1339         return err;
1340 }
1341
1342 int ocfs2_permission(struct inode *inode, int mask)
1343 {
1344         int ret, had_lock;
1345         struct ocfs2_lock_holder oh;
1346
1347         if (mask & MAY_NOT_BLOCK)
1348                 return -ECHILD;
1349
1350         had_lock = ocfs2_inode_lock_tracker(inode, NULL, 0, &oh);
1351         if (had_lock < 0) {
1352                 ret = had_lock;
1353                 goto out;
1354         } else if (had_lock) {
1355                 /* See comments in ocfs2_setattr() for details.
1356                  * The call chain of this case could be:
1357                  * do_sys_open()
1358                  *  may_open()
1359                  *   inode_permission()
1360                  *    ocfs2_permission()
1361                  *     ocfs2_iop_get_acl()
1362                  */
1363                 mlog(ML_ERROR, "Another case of recursive locking:\n");
1364                 dump_stack();
1365         }
1366
1367         ret = generic_permission(inode, mask);
1368
1369         ocfs2_inode_unlock_tracker(inode, 0, &oh, had_lock);
1370 out:
1371         return ret;
1372 }
1373
1374 static int __ocfs2_write_remove_suid(struct inode *inode,
1375                                      struct buffer_head *bh)
1376 {
1377         int ret;
1378         handle_t *handle;
1379         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1380         struct ocfs2_dinode *di;
1381
1382         trace_ocfs2_write_remove_suid(
1383                         (unsigned long long)OCFS2_I(inode)->ip_blkno,
1384                         inode->i_mode);
1385
1386         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1387         if (IS_ERR(handle)) {
1388                 ret = PTR_ERR(handle);
1389                 mlog_errno(ret);
1390                 goto out;
1391         }
1392
1393         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), bh,
1394                                       OCFS2_JOURNAL_ACCESS_WRITE);
1395         if (ret < 0) {
1396                 mlog_errno(ret);
1397                 goto out_trans;
1398         }
1399
1400         inode->i_mode &= ~S_ISUID;
1401         if ((inode->i_mode & S_ISGID) && (inode->i_mode & S_IXGRP))
1402                 inode->i_mode &= ~S_ISGID;
1403
1404         di = (struct ocfs2_dinode *) bh->b_data;
1405         di->i_mode = cpu_to_le16(inode->i_mode);
1406         ocfs2_update_inode_fsync_trans(handle, inode, 0);
1407
1408         ocfs2_journal_dirty(handle, bh);
1409
1410 out_trans:
1411         ocfs2_commit_trans(osb, handle);
1412 out:
1413         return ret;
1414 }
1415
1416 static int ocfs2_write_remove_suid(struct inode *inode)
1417 {
1418         int ret;
1419         struct buffer_head *bh = NULL;
1420
1421         ret = ocfs2_read_inode_block(inode, &bh);
1422         if (ret < 0) {
1423                 mlog_errno(ret);
1424                 goto out;
1425         }
1426
1427         ret =  __ocfs2_write_remove_suid(inode, bh);
1428 out:
1429         brelse(bh);
1430         return ret;
1431 }
1432
1433 /*
1434  * Allocate enough extents to cover the region starting at byte offset
1435  * start for len bytes. Existing extents are skipped, any extents
1436  * added are marked as "unwritten".
1437  */
1438 static int ocfs2_allocate_unwritten_extents(struct inode *inode,
1439                                             u64 start, u64 len)
1440 {
1441         int ret;
1442         u32 cpos, phys_cpos, clusters, alloc_size;
1443         u64 end = start + len;
1444         struct buffer_head *di_bh = NULL;
1445
1446         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1447                 ret = ocfs2_read_inode_block(inode, &di_bh);
1448                 if (ret) {
1449                         mlog_errno(ret);
1450                         goto out;
1451                 }
1452
1453                 /*
1454                  * Nothing to do if the requested reservation range
1455                  * fits within the inode.
1456                  */
1457                 if (ocfs2_size_fits_inline_data(di_bh, end))
1458                         goto out;
1459
1460                 ret = ocfs2_convert_inline_data_to_extents(inode, di_bh);
1461                 if (ret) {
1462                         mlog_errno(ret);
1463                         goto out;
1464                 }
1465         }
1466
1467         /*
1468          * We consider both start and len to be inclusive.
1469          */
1470         cpos = start >> OCFS2_SB(inode->i_sb)->s_clustersize_bits;
1471         clusters = ocfs2_clusters_for_bytes(inode->i_sb, start + len);
1472         clusters -= cpos;
1473
1474         while (clusters) {
1475                 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos,
1476                                          &alloc_size, NULL);
1477                 if (ret) {
1478                         mlog_errno(ret);
1479                         goto out;
1480                 }
1481
1482                 /*
1483                  * Hole or existing extent len can be arbitrary, so
1484                  * cap it to our own allocation request.
1485                  */
1486                 if (alloc_size > clusters)
1487                         alloc_size = clusters;
1488
1489                 if (phys_cpos) {
1490                         /*
1491                          * We already have an allocation at this
1492                          * region so we can safely skip it.
1493                          */
1494                         goto next;
1495                 }
1496
1497                 ret = __ocfs2_extend_allocation(inode, cpos, alloc_size, 1);
1498                 if (ret) {
1499                         if (ret != -ENOSPC)
1500                                 mlog_errno(ret);
1501                         goto out;
1502                 }
1503
1504 next:
1505                 cpos += alloc_size;
1506                 clusters -= alloc_size;
1507         }
1508
1509         ret = 0;
1510 out:
1511
1512         brelse(di_bh);
1513         return ret;
1514 }
1515
1516 /*
1517  * Truncate a byte range, avoiding pages within partial clusters. This
1518  * preserves those pages for the zeroing code to write to.
1519  */
1520 static void ocfs2_truncate_cluster_pages(struct inode *inode, u64 byte_start,
1521                                          u64 byte_len)
1522 {
1523         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1524         loff_t start, end;
1525         struct address_space *mapping = inode->i_mapping;
1526
1527         start = (loff_t)ocfs2_align_bytes_to_clusters(inode->i_sb, byte_start);
1528         end = byte_start + byte_len;
1529         end = end & ~(osb->s_clustersize - 1);
1530
1531         if (start < end) {
1532                 unmap_mapping_range(mapping, start, end - start, 0);
1533                 truncate_inode_pages_range(mapping, start, end - 1);
1534         }
1535 }
1536
1537 static int ocfs2_zero_partial_clusters(struct inode *inode,
1538                                        u64 start, u64 len)
1539 {
1540         int ret = 0;
1541         u64 tmpend = 0;
1542         u64 end = start + len;
1543         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1544         unsigned int csize = osb->s_clustersize;
1545         handle_t *handle;
1546
1547         /*
1548          * The "start" and "end" values are NOT necessarily part of
1549          * the range whose allocation is being deleted. Rather, this
1550          * is what the user passed in with the request. We must zero
1551          * partial clusters here. There's no need to worry about
1552          * physical allocation - the zeroing code knows to skip holes.
1553          */
1554         trace_ocfs2_zero_partial_clusters(
1555                 (unsigned long long)OCFS2_I(inode)->ip_blkno,
1556                 (unsigned long long)start, (unsigned long long)end);
1557
1558         /*
1559          * If both edges are on a cluster boundary then there's no
1560          * zeroing required as the region is part of the allocation to
1561          * be truncated.
1562          */
1563         if ((start & (csize - 1)) == 0 && (end & (csize - 1)) == 0)
1564                 goto out;
1565
1566         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1567         if (IS_ERR(handle)) {
1568                 ret = PTR_ERR(handle);
1569                 mlog_errno(ret);
1570                 goto out;
1571         }
1572
1573         /*
1574          * If start is on a cluster boundary and end is somewhere in another
1575          * cluster, we have not COWed the cluster starting at start, unless
1576          * end is also within the same cluster. So, in this case, we skip this
1577          * first call to ocfs2_zero_range_for_truncate() truncate and move on
1578          * to the next one.
1579          */
1580         if ((start & (csize - 1)) != 0) {
1581                 /*
1582                  * We want to get the byte offset of the end of the 1st
1583                  * cluster.
1584                  */
1585                 tmpend = (u64)osb->s_clustersize +
1586                         (start & ~(osb->s_clustersize - 1));
1587                 if (tmpend > end)
1588                         tmpend = end;
1589
1590                 trace_ocfs2_zero_partial_clusters_range1(
1591                         (unsigned long long)start,
1592                         (unsigned long long)tmpend);
1593
1594                 ret = ocfs2_zero_range_for_truncate(inode, handle, start,
1595                                                     tmpend);
1596                 if (ret)
1597                         mlog_errno(ret);
1598         }
1599
1600         if (tmpend < end) {
1601                 /*
1602                  * This may make start and end equal, but the zeroing
1603                  * code will skip any work in that case so there's no
1604                  * need to catch it up here.
1605                  */
1606                 start = end & ~(osb->s_clustersize - 1);
1607
1608                 trace_ocfs2_zero_partial_clusters_range2(
1609                         (unsigned long long)start, (unsigned long long)end);
1610
1611                 ret = ocfs2_zero_range_for_truncate(inode, handle, start, end);
1612                 if (ret)
1613                         mlog_errno(ret);
1614         }
1615         ocfs2_update_inode_fsync_trans(handle, inode, 1);
1616
1617         ocfs2_commit_trans(osb, handle);
1618 out:
1619         return ret;
1620 }
1621
1622 static int ocfs2_find_rec(struct ocfs2_extent_list *el, u32 pos)
1623 {
1624         int i;
1625         struct ocfs2_extent_rec *rec = NULL;
1626
1627         for (i = le16_to_cpu(el->l_next_free_rec) - 1; i >= 0; i--) {
1628
1629                 rec = &el->l_recs[i];
1630
1631                 if (le32_to_cpu(rec->e_cpos) < pos)
1632                         break;
1633         }
1634
1635         return i;
1636 }
1637
1638 /*
1639  * Helper to calculate the punching pos and length in one run, we handle the
1640  * following three cases in order:
1641  *
1642  * - remove the entire record
1643  * - remove a partial record
1644  * - no record needs to be removed (hole-punching completed)
1645 */
1646 static void ocfs2_calc_trunc_pos(struct inode *inode,
1647                                  struct ocfs2_extent_list *el,
1648                                  struct ocfs2_extent_rec *rec,
1649                                  u32 trunc_start, u32 *trunc_cpos,
1650                                  u32 *trunc_len, u32 *trunc_end,
1651                                  u64 *blkno, int *done)
1652 {
1653         int ret = 0;
1654         u32 coff, range;
1655
1656         range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
1657
1658         if (le32_to_cpu(rec->e_cpos) >= trunc_start) {
1659                 /*
1660                  * remove an entire extent record.
1661                  */
1662                 *trunc_cpos = le32_to_cpu(rec->e_cpos);
1663                 /*
1664                  * Skip holes if any.
1665                  */
1666                 if (range < *trunc_end)
1667                         *trunc_end = range;
1668                 *trunc_len = *trunc_end - le32_to_cpu(rec->e_cpos);
1669                 *blkno = le64_to_cpu(rec->e_blkno);
1670                 *trunc_end = le32_to_cpu(rec->e_cpos);
1671         } else if (range > trunc_start) {
1672                 /*
1673                  * remove a partial extent record, which means we're
1674                  * removing the last extent record.
1675                  */
1676                 *trunc_cpos = trunc_start;
1677                 /*
1678                  * skip hole if any.
1679                  */
1680                 if (range < *trunc_end)
1681                         *trunc_end = range;
1682                 *trunc_len = *trunc_end - trunc_start;
1683                 coff = trunc_start - le32_to_cpu(rec->e_cpos);
1684                 *blkno = le64_to_cpu(rec->e_blkno) +
1685                                 ocfs2_clusters_to_blocks(inode->i_sb, coff);
1686                 *trunc_end = trunc_start;
1687         } else {
1688                 /*
1689                  * It may have two following possibilities:
1690                  *
1691                  * - last record has been removed
1692                  * - trunc_start was within a hole
1693                  *
1694                  * both two cases mean the completion of hole punching.
1695                  */
1696                 ret = 1;
1697         }
1698
1699         *done = ret;
1700 }
1701
1702 int ocfs2_remove_inode_range(struct inode *inode,
1703                              struct buffer_head *di_bh, u64 byte_start,
1704                              u64 byte_len)
1705 {
1706         int ret = 0, flags = 0, done = 0, i;
1707         u32 trunc_start, trunc_len, trunc_end, trunc_cpos, phys_cpos;
1708         u32 cluster_in_el;
1709         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1710         struct ocfs2_cached_dealloc_ctxt dealloc;
1711         struct address_space *mapping = inode->i_mapping;
1712         struct ocfs2_extent_tree et;
1713         struct ocfs2_path *path = NULL;
1714         struct ocfs2_extent_list *el = NULL;
1715         struct ocfs2_extent_rec *rec = NULL;
1716         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
1717         u64 blkno, refcount_loc = le64_to_cpu(di->i_refcount_loc);
1718
1719         ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), di_bh);
1720         ocfs2_init_dealloc_ctxt(&dealloc);
1721
1722         trace_ocfs2_remove_inode_range(
1723                         (unsigned long long)OCFS2_I(inode)->ip_blkno,
1724                         (unsigned long long)byte_start,
1725                         (unsigned long long)byte_len);
1726
1727         if (byte_len == 0)
1728                 return 0;
1729
1730         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1731                 ret = ocfs2_truncate_inline(inode, di_bh, byte_start,
1732                                             byte_start + byte_len, 0);
1733                 if (ret) {
1734                         mlog_errno(ret);
1735                         goto out;
1736                 }
1737                 /*
1738                  * There's no need to get fancy with the page cache
1739                  * truncate of an inline-data inode. We're talking
1740                  * about less than a page here, which will be cached
1741                  * in the dinode buffer anyway.
1742                  */
1743                 unmap_mapping_range(mapping, 0, 0, 0);
1744                 truncate_inode_pages(mapping, 0);
1745                 goto out;
1746         }
1747
1748         /*
1749          * For reflinks, we may need to CoW 2 clusters which might be
1750          * partially zero'd later, if hole's start and end offset were
1751          * within one cluster(means is not exactly aligned to clustersize).
1752          */
1753
1754         if (ocfs2_is_refcount_inode(inode)) {
1755                 ret = ocfs2_cow_file_pos(inode, di_bh, byte_start);
1756                 if (ret) {
1757                         mlog_errno(ret);
1758                         goto out;
1759                 }
1760
1761                 ret = ocfs2_cow_file_pos(inode, di_bh, byte_start + byte_len);
1762                 if (ret) {
1763                         mlog_errno(ret);
1764                         goto out;
1765                 }
1766         }
1767
1768         trunc_start = ocfs2_clusters_for_bytes(osb->sb, byte_start);
1769         trunc_end = (byte_start + byte_len) >> osb->s_clustersize_bits;
1770         cluster_in_el = trunc_end;
1771
1772         ret = ocfs2_zero_partial_clusters(inode, byte_start, byte_len);
1773         if (ret) {
1774                 mlog_errno(ret);
1775                 goto out;
1776         }
1777
1778         path = ocfs2_new_path_from_et(&et);
1779         if (!path) {
1780                 ret = -ENOMEM;
1781                 mlog_errno(ret);
1782                 goto out;
1783         }
1784
1785         while (trunc_end > trunc_start) {
1786
1787                 ret = ocfs2_find_path(INODE_CACHE(inode), path,
1788                                       cluster_in_el);
1789                 if (ret) {
1790                         mlog_errno(ret);
1791                         goto out;
1792                 }
1793
1794                 el = path_leaf_el(path);
1795
1796                 i = ocfs2_find_rec(el, trunc_end);
1797                 /*
1798                  * Need to go to previous extent block.
1799                  */
1800                 if (i < 0) {
1801                         if (path->p_tree_depth == 0)
1802                                 break;
1803
1804                         ret = ocfs2_find_cpos_for_left_leaf(inode->i_sb,
1805                                                             path,
1806                                                             &cluster_in_el);
1807                         if (ret) {
1808                                 mlog_errno(ret);
1809                                 goto out;
1810                         }
1811
1812                         /*
1813                          * We've reached the leftmost extent block,
1814                          * it's safe to leave.
1815                          */
1816                         if (cluster_in_el == 0)
1817                                 break;
1818
1819                         /*
1820                          * The 'pos' searched for previous extent block is
1821                          * always one cluster less than actual trunc_end.
1822                          */
1823                         trunc_end = cluster_in_el + 1;
1824
1825                         ocfs2_reinit_path(path, 1);
1826
1827                         continue;
1828
1829                 } else
1830                         rec = &el->l_recs[i];
1831
1832                 ocfs2_calc_trunc_pos(inode, el, rec, trunc_start, &trunc_cpos,
1833                                      &trunc_len, &trunc_end, &blkno, &done);
1834                 if (done)
1835                         break;
1836
1837                 flags = rec->e_flags;
1838                 phys_cpos = ocfs2_blocks_to_clusters(inode->i_sb, blkno);
1839
1840                 ret = ocfs2_remove_btree_range(inode, &et, trunc_cpos,
1841                                                phys_cpos, trunc_len, flags,
1842                                                &dealloc, refcount_loc, false);
1843                 if (ret < 0) {
1844                         mlog_errno(ret);
1845                         goto out;
1846                 }
1847
1848                 cluster_in_el = trunc_end;
1849
1850                 ocfs2_reinit_path(path, 1);
1851         }
1852
1853         ocfs2_truncate_cluster_pages(inode, byte_start, byte_len);
1854
1855 out:
1856         ocfs2_free_path(path);
1857         ocfs2_schedule_truncate_log_flush(osb, 1);
1858         ocfs2_run_deallocs(osb, &dealloc);
1859
1860         return ret;
1861 }
1862
1863 /*
1864  * Parts of this function taken from xfs_change_file_space()
1865  */
1866 static int __ocfs2_change_file_space(struct file *file, struct inode *inode,
1867                                      loff_t f_pos, unsigned int cmd,
1868                                      struct ocfs2_space_resv *sr,
1869                                      int change_size)
1870 {
1871         int ret;
1872         s64 llen;
1873         loff_t size;
1874         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1875         struct buffer_head *di_bh = NULL;
1876         handle_t *handle;
1877         unsigned long long max_off = inode->i_sb->s_maxbytes;
1878
1879         if (ocfs2_is_hard_readonly(osb) || ocfs2_is_soft_readonly(osb))
1880                 return -EROFS;
1881
1882         inode_lock(inode);
1883
1884         /*
1885          * This prevents concurrent writes on other nodes
1886          */
1887         ret = ocfs2_rw_lock(inode, 1);
1888         if (ret) {
1889                 mlog_errno(ret);
1890                 goto out;
1891         }
1892
1893         ret = ocfs2_inode_lock(inode, &di_bh, 1);
1894         if (ret) {
1895                 mlog_errno(ret);
1896                 goto out_rw_unlock;
1897         }
1898
1899         if (inode->i_flags & (S_IMMUTABLE|S_APPEND)) {
1900                 ret = -EPERM;
1901                 goto out_inode_unlock;
1902         }
1903
1904         switch (sr->l_whence) {
1905         case 0: /*SEEK_SET*/
1906                 break;
1907         case 1: /*SEEK_CUR*/
1908                 sr->l_start += f_pos;
1909                 break;
1910         case 2: /*SEEK_END*/
1911                 sr->l_start += i_size_read(inode);
1912                 break;
1913         default:
1914                 ret = -EINVAL;
1915                 goto out_inode_unlock;
1916         }
1917         sr->l_whence = 0;
1918
1919         llen = sr->l_len > 0 ? sr->l_len - 1 : sr->l_len;
1920
1921         if (sr->l_start < 0
1922             || sr->l_start > max_off
1923             || (sr->l_start + llen) < 0
1924             || (sr->l_start + llen) > max_off) {
1925                 ret = -EINVAL;
1926                 goto out_inode_unlock;
1927         }
1928         size = sr->l_start + sr->l_len;
1929
1930         if (cmd == OCFS2_IOC_RESVSP || cmd == OCFS2_IOC_RESVSP64 ||
1931             cmd == OCFS2_IOC_UNRESVSP || cmd == OCFS2_IOC_UNRESVSP64) {
1932                 if (sr->l_len <= 0) {
1933                         ret = -EINVAL;
1934                         goto out_inode_unlock;
1935                 }
1936         }
1937
1938         if (file && should_remove_suid(file->f_path.dentry)) {
1939                 ret = __ocfs2_write_remove_suid(inode, di_bh);
1940                 if (ret) {
1941                         mlog_errno(ret);
1942                         goto out_inode_unlock;
1943                 }
1944         }
1945
1946         down_write(&OCFS2_I(inode)->ip_alloc_sem);
1947         switch (cmd) {
1948         case OCFS2_IOC_RESVSP:
1949         case OCFS2_IOC_RESVSP64:
1950                 /*
1951                  * This takes unsigned offsets, but the signed ones we
1952                  * pass have been checked against overflow above.
1953                  */
1954                 ret = ocfs2_allocate_unwritten_extents(inode, sr->l_start,
1955                                                        sr->l_len);
1956                 break;
1957         case OCFS2_IOC_UNRESVSP:
1958         case OCFS2_IOC_UNRESVSP64:
1959                 ret = ocfs2_remove_inode_range(inode, di_bh, sr->l_start,
1960                                                sr->l_len);
1961                 break;
1962         default:
1963                 ret = -EINVAL;
1964         }
1965         up_write(&OCFS2_I(inode)->ip_alloc_sem);
1966         if (ret) {
1967                 mlog_errno(ret);
1968                 goto out_inode_unlock;
1969         }
1970
1971         /*
1972          * We update c/mtime for these changes
1973          */
1974         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1975         if (IS_ERR(handle)) {
1976                 ret = PTR_ERR(handle);
1977                 mlog_errno(ret);
1978                 goto out_inode_unlock;
1979         }
1980
1981         if (change_size && i_size_read(inode) < size)
1982                 i_size_write(inode, size);
1983
1984         inode->i_ctime = inode->i_mtime = current_time(inode);
1985         ret = ocfs2_mark_inode_dirty(handle, inode, di_bh);
1986         if (ret < 0)
1987                 mlog_errno(ret);
1988
1989         if (file && (file->f_flags & O_SYNC))
1990                 handle->h_sync = 1;
1991
1992         ocfs2_commit_trans(osb, handle);
1993
1994 out_inode_unlock:
1995         brelse(di_bh);
1996         ocfs2_inode_unlock(inode, 1);
1997 out_rw_unlock:
1998         ocfs2_rw_unlock(inode, 1);
1999
2000 out:
2001         inode_unlock(inode);
2002         return ret;
2003 }
2004
2005 int ocfs2_change_file_space(struct file *file, unsigned int cmd,
2006                             struct ocfs2_space_resv *sr)
2007 {
2008         struct inode *inode = file_inode(file);
2009         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
2010         int ret;
2011
2012         if ((cmd == OCFS2_IOC_RESVSP || cmd == OCFS2_IOC_RESVSP64) &&
2013             !ocfs2_writes_unwritten_extents(osb))
2014                 return -ENOTTY;
2015         else if ((cmd == OCFS2_IOC_UNRESVSP || cmd == OCFS2_IOC_UNRESVSP64) &&
2016                  !ocfs2_sparse_alloc(osb))
2017                 return -ENOTTY;
2018
2019         if (!S_ISREG(inode->i_mode))
2020                 return -EINVAL;
2021
2022         if (!(file->f_mode & FMODE_WRITE))
2023                 return -EBADF;
2024
2025         ret = mnt_want_write_file(file);
2026         if (ret)
2027                 return ret;
2028         ret = __ocfs2_change_file_space(file, inode, file->f_pos, cmd, sr, 0);
2029         mnt_drop_write_file(file);
2030         return ret;
2031 }
2032
2033 static long ocfs2_fallocate(struct file *file, int mode, loff_t offset,
2034                             loff_t len)
2035 {
2036         struct inode *inode = file_inode(file);
2037         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
2038         struct ocfs2_space_resv sr;
2039         int change_size = 1;
2040         int cmd = OCFS2_IOC_RESVSP64;
2041
2042         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
2043                 return -EOPNOTSUPP;
2044         if (!ocfs2_writes_unwritten_extents(osb))
2045                 return -EOPNOTSUPP;
2046
2047         if (mode & FALLOC_FL_KEEP_SIZE)
2048                 change_size = 0;
2049
2050         if (mode & FALLOC_FL_PUNCH_HOLE)
2051                 cmd = OCFS2_IOC_UNRESVSP64;
2052
2053         sr.l_whence = 0;
2054         sr.l_start = (s64)offset;
2055         sr.l_len = (s64)len;
2056
2057         return __ocfs2_change_file_space(NULL, inode, offset, cmd, &sr,
2058                                          change_size);
2059 }
2060
2061 int ocfs2_check_range_for_refcount(struct inode *inode, loff_t pos,
2062                                    size_t count)
2063 {
2064         int ret = 0;
2065         unsigned int extent_flags;
2066         u32 cpos, clusters, extent_len, phys_cpos;
2067         struct super_block *sb = inode->i_sb;
2068
2069         if (!ocfs2_refcount_tree(OCFS2_SB(inode->i_sb)) ||
2070             !ocfs2_is_refcount_inode(inode) ||
2071             OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
2072                 return 0;
2073
2074         cpos = pos >> OCFS2_SB(sb)->s_clustersize_bits;
2075         clusters = ocfs2_clusters_for_bytes(sb, pos + count) - cpos;
2076
2077         while (clusters) {
2078                 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos, &extent_len,
2079                                          &extent_flags);
2080                 if (ret < 0) {
2081                         mlog_errno(ret);
2082                         goto out;
2083                 }
2084
2085                 if (phys_cpos && (extent_flags & OCFS2_EXT_REFCOUNTED)) {
2086                         ret = 1;
2087                         break;
2088                 }
2089
2090                 if (extent_len > clusters)
2091                         extent_len = clusters;
2092
2093                 clusters -= extent_len;
2094                 cpos += extent_len;
2095         }
2096 out:
2097         return ret;
2098 }
2099
2100 static int ocfs2_is_io_unaligned(struct inode *inode, size_t count, loff_t pos)
2101 {
2102         int blockmask = inode->i_sb->s_blocksize - 1;
2103         loff_t final_size = pos + count;
2104
2105         if ((pos & blockmask) || (final_size & blockmask))
2106                 return 1;
2107         return 0;
2108 }
2109
2110 static int ocfs2_prepare_inode_for_refcount(struct inode *inode,
2111                                             struct file *file,
2112                                             loff_t pos, size_t count,
2113                                             int *meta_level)
2114 {
2115         int ret;
2116         struct buffer_head *di_bh = NULL;
2117         u32 cpos = pos >> OCFS2_SB(inode->i_sb)->s_clustersize_bits;
2118         u32 clusters =
2119                 ocfs2_clusters_for_bytes(inode->i_sb, pos + count) - cpos;
2120
2121         ret = ocfs2_inode_lock(inode, &di_bh, 1);
2122         if (ret) {
2123                 mlog_errno(ret);
2124                 goto out;
2125         }
2126
2127         *meta_level = 1;
2128
2129         ret = ocfs2_refcount_cow(inode, di_bh, cpos, clusters, UINT_MAX);
2130         if (ret)
2131                 mlog_errno(ret);
2132 out:
2133         brelse(di_bh);
2134         return ret;
2135 }
2136
2137 static int ocfs2_prepare_inode_for_write(struct file *file,
2138                                          loff_t pos,
2139                                          size_t count)
2140 {
2141         int ret = 0, meta_level = 0;
2142         struct dentry *dentry = file->f_path.dentry;
2143         struct inode *inode = d_inode(dentry);
2144         loff_t end;
2145
2146         /*
2147          * We start with a read level meta lock and only jump to an ex
2148          * if we need to make modifications here.
2149          */
2150         for(;;) {
2151                 ret = ocfs2_inode_lock(inode, NULL, meta_level);
2152                 if (ret < 0) {
2153                         meta_level = -1;
2154                         mlog_errno(ret);
2155                         goto out;
2156                 }
2157
2158                 /* Clear suid / sgid if necessary. We do this here
2159                  * instead of later in the write path because
2160                  * remove_suid() calls ->setattr without any hint that
2161                  * we may have already done our cluster locking. Since
2162                  * ocfs2_setattr() *must* take cluster locks to
2163                  * proceed, this will lead us to recursively lock the
2164                  * inode. There's also the dinode i_size state which
2165                  * can be lost via setattr during extending writes (we
2166                  * set inode->i_size at the end of a write. */
2167                 if (should_remove_suid(dentry)) {
2168                         if (meta_level == 0) {
2169                                 ocfs2_inode_unlock(inode, meta_level);
2170                                 meta_level = 1;
2171                                 continue;
2172                         }
2173
2174                         ret = ocfs2_write_remove_suid(inode);
2175                         if (ret < 0) {
2176                                 mlog_errno(ret);
2177                                 goto out_unlock;
2178                         }
2179                 }
2180
2181                 end = pos + count;
2182
2183                 ret = ocfs2_check_range_for_refcount(inode, pos, count);
2184                 if (ret == 1) {
2185                         ocfs2_inode_unlock(inode, meta_level);
2186                         meta_level = -1;
2187
2188                         ret = ocfs2_prepare_inode_for_refcount(inode,
2189                                                                file,
2190                                                                pos,
2191                                                                count,
2192                                                                &meta_level);
2193                 }
2194
2195                 if (ret < 0) {
2196                         mlog_errno(ret);
2197                         goto out_unlock;
2198                 }
2199
2200                 break;
2201         }
2202
2203 out_unlock:
2204         trace_ocfs2_prepare_inode_for_write(OCFS2_I(inode)->ip_blkno,
2205                                             pos, count);
2206
2207         if (meta_level >= 0)
2208                 ocfs2_inode_unlock(inode, meta_level);
2209
2210 out:
2211         return ret;
2212 }
2213
2214 static ssize_t ocfs2_file_write_iter(struct kiocb *iocb,
2215                                     struct iov_iter *from)
2216 {
2217         int direct_io, rw_level;
2218         ssize_t written = 0;
2219         ssize_t ret;
2220         size_t count = iov_iter_count(from);
2221         struct file *file = iocb->ki_filp;
2222         struct inode *inode = file_inode(file);
2223         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
2224         int full_coherency = !(osb->s_mount_opt &
2225                                OCFS2_MOUNT_COHERENCY_BUFFERED);
2226         void *saved_ki_complete = NULL;
2227         int append_write = ((iocb->ki_pos + count) >=
2228                         i_size_read(inode) ? 1 : 0);
2229
2230         trace_ocfs2_file_aio_write(inode, file, file->f_path.dentry,
2231                 (unsigned long long)OCFS2_I(inode)->ip_blkno,
2232                 file->f_path.dentry->d_name.len,
2233                 file->f_path.dentry->d_name.name,
2234                 (unsigned int)from->nr_segs);   /* GRRRRR */
2235
2236         if (count == 0)
2237                 return 0;
2238
2239         direct_io = iocb->ki_flags & IOCB_DIRECT ? 1 : 0;
2240
2241         inode_lock(inode);
2242
2243         /*
2244          * Concurrent O_DIRECT writes are allowed with
2245          * mount_option "coherency=buffered".
2246          * For append write, we must take rw EX.
2247          */
2248         rw_level = (!direct_io || full_coherency || append_write);
2249
2250         ret = ocfs2_rw_lock(inode, rw_level);
2251         if (ret < 0) {
2252                 mlog_errno(ret);
2253                 goto out_mutex;
2254         }
2255
2256         /*
2257          * O_DIRECT writes with "coherency=full" need to take EX cluster
2258          * inode_lock to guarantee coherency.
2259          */
2260         if (direct_io && full_coherency) {
2261                 /*
2262                  * We need to take and drop the inode lock to force
2263                  * other nodes to drop their caches.  Buffered I/O
2264                  * already does this in write_begin().
2265                  */
2266                 ret = ocfs2_inode_lock(inode, NULL, 1);
2267                 if (ret < 0) {
2268                         mlog_errno(ret);
2269                         goto out;
2270                 }
2271
2272                 ocfs2_inode_unlock(inode, 1);
2273         }
2274
2275         ret = generic_write_checks(iocb, from);
2276         if (ret <= 0) {
2277                 if (ret)
2278                         mlog_errno(ret);
2279                 goto out;
2280         }
2281         count = ret;
2282
2283         ret = ocfs2_prepare_inode_for_write(file, iocb->ki_pos, count);
2284         if (ret < 0) {
2285                 mlog_errno(ret);
2286                 goto out;
2287         }
2288
2289         if (direct_io && !is_sync_kiocb(iocb) &&
2290             ocfs2_is_io_unaligned(inode, count, iocb->ki_pos)) {
2291                 /*
2292                  * Make it a sync io if it's an unaligned aio.
2293                  */
2294                 saved_ki_complete = xchg(&iocb->ki_complete, NULL);
2295         }
2296
2297         /* communicate with ocfs2_dio_end_io */
2298         ocfs2_iocb_set_rw_locked(iocb, rw_level);
2299
2300         written = __generic_file_write_iter(iocb, from);
2301         /* buffered aio wouldn't have proper lock coverage today */
2302         BUG_ON(written == -EIOCBQUEUED && !(iocb->ki_flags & IOCB_DIRECT));
2303
2304         /*
2305          * deep in g_f_a_w_n()->ocfs2_direct_IO we pass in a ocfs2_dio_end_io
2306          * function pointer which is called when o_direct io completes so that
2307          * it can unlock our rw lock.
2308          * Unfortunately there are error cases which call end_io and others
2309          * that don't.  so we don't have to unlock the rw_lock if either an
2310          * async dio is going to do it in the future or an end_io after an
2311          * error has already done it.
2312          */
2313         if ((written == -EIOCBQUEUED) || (!ocfs2_iocb_is_rw_locked(iocb))) {
2314                 rw_level = -1;
2315         }
2316
2317         if (unlikely(written <= 0))
2318                 goto out;
2319
2320         if (((file->f_flags & O_DSYNC) && !direct_io) ||
2321             IS_SYNC(inode)) {
2322                 ret = filemap_fdatawrite_range(file->f_mapping,
2323                                                iocb->ki_pos - written,
2324                                                iocb->ki_pos - 1);
2325                 if (ret < 0)
2326                         written = ret;
2327
2328                 if (!ret) {
2329                         ret = jbd2_journal_force_commit(osb->journal->j_journal);
2330                         if (ret < 0)
2331                                 written = ret;
2332                 }
2333
2334                 if (!ret)
2335                         ret = filemap_fdatawait_range(file->f_mapping,
2336                                                       iocb->ki_pos - written,
2337                                                       iocb->ki_pos - 1);
2338         }
2339
2340 out:
2341         if (saved_ki_complete)
2342                 xchg(&iocb->ki_complete, saved_ki_complete);
2343
2344         if (rw_level != -1)
2345                 ocfs2_rw_unlock(inode, rw_level);
2346
2347 out_mutex:
2348         inode_unlock(inode);
2349
2350         if (written)
2351                 ret = written;
2352         return ret;
2353 }
2354
2355 static ssize_t ocfs2_file_read_iter(struct kiocb *iocb,
2356                                    struct iov_iter *to)
2357 {
2358         int ret = 0, rw_level = -1, lock_level = 0;
2359         struct file *filp = iocb->ki_filp;
2360         struct inode *inode = file_inode(filp);
2361
2362         trace_ocfs2_file_aio_read(inode, filp, filp->f_path.dentry,
2363                         (unsigned long long)OCFS2_I(inode)->ip_blkno,
2364                         filp->f_path.dentry->d_name.len,
2365                         filp->f_path.dentry->d_name.name,
2366                         to->nr_segs);   /* GRRRRR */
2367
2368
2369         if (!inode) {
2370                 ret = -EINVAL;
2371                 mlog_errno(ret);
2372                 goto bail;
2373         }
2374
2375         /*
2376          * buffered reads protect themselves in ->readpage().  O_DIRECT reads
2377          * need locks to protect pending reads from racing with truncate.
2378          */
2379         if (iocb->ki_flags & IOCB_DIRECT) {
2380                 ret = ocfs2_rw_lock(inode, 0);
2381                 if (ret < 0) {
2382                         mlog_errno(ret);
2383                         goto bail;
2384                 }
2385                 rw_level = 0;
2386                 /* communicate with ocfs2_dio_end_io */
2387                 ocfs2_iocb_set_rw_locked(iocb, rw_level);
2388         }
2389
2390         /*
2391          * We're fine letting folks race truncates and extending
2392          * writes with read across the cluster, just like they can
2393          * locally. Hence no rw_lock during read.
2394          *
2395          * Take and drop the meta data lock to update inode fields
2396          * like i_size. This allows the checks down below
2397          * generic_file_aio_read() a chance of actually working.
2398          */
2399         ret = ocfs2_inode_lock_atime(inode, filp->f_path.mnt, &lock_level);
2400         if (ret < 0) {
2401                 mlog_errno(ret);
2402                 goto bail;
2403         }
2404         ocfs2_inode_unlock(inode, lock_level);
2405
2406         ret = generic_file_read_iter(iocb, to);
2407         trace_generic_file_aio_read_ret(ret);
2408
2409         /* buffered aio wouldn't have proper lock coverage today */
2410         BUG_ON(ret == -EIOCBQUEUED && !(iocb->ki_flags & IOCB_DIRECT));
2411
2412         /* see ocfs2_file_write_iter */
2413         if (ret == -EIOCBQUEUED || !ocfs2_iocb_is_rw_locked(iocb)) {
2414                 rw_level = -1;
2415         }
2416
2417 bail:
2418         if (rw_level != -1)
2419                 ocfs2_rw_unlock(inode, rw_level);
2420
2421         return ret;
2422 }
2423
2424 /* Refer generic_file_llseek_unlocked() */
2425 static loff_t ocfs2_file_llseek(struct file *file, loff_t offset, int whence)
2426 {
2427         struct inode *inode = file->f_mapping->host;
2428         int ret = 0;
2429
2430         inode_lock(inode);
2431
2432         switch (whence) {
2433         case SEEK_SET:
2434                 break;
2435         case SEEK_END:
2436                 /* SEEK_END requires the OCFS2 inode lock for the file
2437                  * because it references the file's size.
2438                  */
2439                 ret = ocfs2_inode_lock(inode, NULL, 0);
2440                 if (ret < 0) {
2441                         mlog_errno(ret);
2442                         goto out;
2443                 }
2444                 offset += i_size_read(inode);
2445                 ocfs2_inode_unlock(inode, 0);
2446                 break;
2447         case SEEK_CUR:
2448                 if (offset == 0) {
2449                         offset = file->f_pos;
2450                         goto out;
2451                 }
2452                 offset += file->f_pos;
2453                 break;
2454         case SEEK_DATA:
2455         case SEEK_HOLE:
2456                 ret = ocfs2_seek_data_hole_offset(file, &offset, whence);
2457                 if (ret)
2458                         goto out;
2459                 break;
2460         default:
2461                 ret = -EINVAL;
2462                 goto out;
2463         }
2464
2465         offset = vfs_setpos(file, offset, inode->i_sb->s_maxbytes);
2466
2467 out:
2468         inode_unlock(inode);
2469         if (ret)
2470                 return ret;
2471         return offset;
2472 }
2473
2474 static int ocfs2_file_clone_range(struct file *file_in,
2475                                   loff_t pos_in,
2476                                   struct file *file_out,
2477                                   loff_t pos_out,
2478                                   u64 len)
2479 {
2480         return ocfs2_reflink_remap_range(file_in, pos_in, file_out, pos_out,
2481                                          len, false);
2482 }
2483
2484 static ssize_t ocfs2_file_dedupe_range(struct file *src_file,
2485                                        u64 loff,
2486                                        u64 len,
2487                                        struct file *dst_file,
2488                                        u64 dst_loff)
2489 {
2490         int error;
2491
2492         error = ocfs2_reflink_remap_range(src_file, loff, dst_file, dst_loff,
2493                                           len, true);
2494         if (error)
2495                 return error;
2496         return len;
2497 }
2498
2499 const struct inode_operations ocfs2_file_iops = {
2500         .setattr        = ocfs2_setattr,
2501         .getattr        = ocfs2_getattr,
2502         .permission     = ocfs2_permission,
2503         .listxattr      = ocfs2_listxattr,
2504         .fiemap         = ocfs2_fiemap,
2505         .get_acl        = ocfs2_iop_get_acl,
2506         .set_acl        = ocfs2_iop_set_acl,
2507 };
2508
2509 const struct inode_operations ocfs2_special_file_iops = {
2510         .setattr        = ocfs2_setattr,
2511         .getattr        = ocfs2_getattr,
2512         .permission     = ocfs2_permission,
2513         .get_acl        = ocfs2_iop_get_acl,
2514         .set_acl        = ocfs2_iop_set_acl,
2515 };
2516
2517 /*
2518  * Other than ->lock, keep ocfs2_fops and ocfs2_dops in sync with
2519  * ocfs2_fops_no_plocks and ocfs2_dops_no_plocks!
2520  */
2521 const struct file_operations ocfs2_fops = {
2522         .llseek         = ocfs2_file_llseek,
2523         .mmap           = ocfs2_mmap,
2524         .fsync          = ocfs2_sync_file,
2525         .release        = ocfs2_file_release,
2526         .open           = ocfs2_file_open,
2527         .read_iter      = ocfs2_file_read_iter,
2528         .write_iter     = ocfs2_file_write_iter,
2529         .unlocked_ioctl = ocfs2_ioctl,
2530 #ifdef CONFIG_COMPAT
2531         .compat_ioctl   = ocfs2_compat_ioctl,
2532 #endif
2533         .lock           = ocfs2_lock,
2534         .flock          = ocfs2_flock,
2535         .splice_read    = generic_file_splice_read,
2536         .splice_write   = iter_file_splice_write,
2537         .fallocate      = ocfs2_fallocate,
2538         .clone_file_range = ocfs2_file_clone_range,
2539         .dedupe_file_range = ocfs2_file_dedupe_range,
2540 };
2541
2542 const struct file_operations ocfs2_dops = {
2543         .llseek         = generic_file_llseek,
2544         .read           = generic_read_dir,
2545         .iterate        = ocfs2_readdir,
2546         .fsync          = ocfs2_sync_file,
2547         .release        = ocfs2_dir_release,
2548         .open           = ocfs2_dir_open,
2549         .unlocked_ioctl = ocfs2_ioctl,
2550 #ifdef CONFIG_COMPAT
2551         .compat_ioctl   = ocfs2_compat_ioctl,
2552 #endif
2553         .lock           = ocfs2_lock,
2554         .flock          = ocfs2_flock,
2555 };
2556
2557 /*
2558  * POSIX-lockless variants of our file_operations.
2559  *
2560  * These will be used if the underlying cluster stack does not support
2561  * posix file locking, if the user passes the "localflocks" mount
2562  * option, or if we have a local-only fs.
2563  *
2564  * ocfs2_flock is in here because all stacks handle UNIX file locks,
2565  * so we still want it in the case of no stack support for
2566  * plocks. Internally, it will do the right thing when asked to ignore
2567  * the cluster.
2568  */
2569 const struct file_operations ocfs2_fops_no_plocks = {
2570         .llseek         = ocfs2_file_llseek,
2571         .mmap           = ocfs2_mmap,
2572         .fsync          = ocfs2_sync_file,
2573         .release        = ocfs2_file_release,
2574         .open           = ocfs2_file_open,
2575         .read_iter      = ocfs2_file_read_iter,
2576         .write_iter     = ocfs2_file_write_iter,
2577         .unlocked_ioctl = ocfs2_ioctl,
2578 #ifdef CONFIG_COMPAT
2579         .compat_ioctl   = ocfs2_compat_ioctl,
2580 #endif
2581         .flock          = ocfs2_flock,
2582         .splice_read    = generic_file_splice_read,
2583         .splice_write   = iter_file_splice_write,
2584         .fallocate      = ocfs2_fallocate,
2585         .clone_file_range = ocfs2_file_clone_range,
2586         .dedupe_file_range = ocfs2_file_dedupe_range,
2587 };
2588
2589 const struct file_operations ocfs2_dops_no_plocks = {
2590         .llseek         = generic_file_llseek,
2591         .read           = generic_read_dir,
2592         .iterate        = ocfs2_readdir,
2593         .fsync          = ocfs2_sync_file,
2594         .release        = ocfs2_dir_release,
2595         .open           = ocfs2_dir_open,
2596         .unlocked_ioctl = ocfs2_ioctl,
2597 #ifdef CONFIG_COMPAT
2598         .compat_ioctl   = ocfs2_compat_ioctl,
2599 #endif
2600         .flock          = ocfs2_flock,
2601 };