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(const struct path *path, struct kstat *stat,
1310                   u32 request_mask, unsigned int flags)
1311 {
1312         struct inode *inode = d_inode(path->dentry);
1313         struct super_block *sb = path->dentry->d_sb;
1314         struct ocfs2_super *osb = sb->s_fs_info;
1315         int err;
1316
1317         err = ocfs2_inode_revalidate(path->dentry);
1318         if (err) {
1319                 if (err != -ENOENT)
1320                         mlog_errno(err);
1321                 goto bail;
1322         }
1323
1324         generic_fillattr(inode, stat);
1325         /*
1326          * If there is inline data in the inode, the inode will normally not
1327          * have data blocks allocated (it may have an external xattr block).
1328          * Report at least one sector for such files, so tools like tar, rsync,
1329          * others don't incorrectly think the file is completely sparse.
1330          */
1331         if (unlikely(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL))
1332                 stat->blocks += (stat->size + 511)>>9;
1333
1334         /* We set the blksize from the cluster size for performance */
1335         stat->blksize = osb->s_clustersize;
1336
1337 bail:
1338         return err;
1339 }
1340
1341 int ocfs2_permission(struct inode *inode, int mask)
1342 {
1343         int ret, had_lock;
1344         struct ocfs2_lock_holder oh;
1345
1346         if (mask & MAY_NOT_BLOCK)
1347                 return -ECHILD;
1348
1349         had_lock = ocfs2_inode_lock_tracker(inode, NULL, 0, &oh);
1350         if (had_lock < 0) {
1351                 ret = had_lock;
1352                 goto out;
1353         } else if (had_lock) {
1354                 /* See comments in ocfs2_setattr() for details.
1355                  * The call chain of this case could be:
1356                  * do_sys_open()
1357                  *  may_open()
1358                  *   inode_permission()
1359                  *    ocfs2_permission()
1360                  *     ocfs2_iop_get_acl()
1361                  */
1362                 mlog(ML_ERROR, "Another case of recursive locking:\n");
1363                 dump_stack();
1364         }
1365
1366         ret = generic_permission(inode, mask);
1367
1368         ocfs2_inode_unlock_tracker(inode, 0, &oh, had_lock);
1369 out:
1370         return ret;
1371 }
1372
1373 static int __ocfs2_write_remove_suid(struct inode *inode,
1374                                      struct buffer_head *bh)
1375 {
1376         int ret;
1377         handle_t *handle;
1378         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1379         struct ocfs2_dinode *di;
1380
1381         trace_ocfs2_write_remove_suid(
1382                         (unsigned long long)OCFS2_I(inode)->ip_blkno,
1383                         inode->i_mode);
1384
1385         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1386         if (IS_ERR(handle)) {
1387                 ret = PTR_ERR(handle);
1388                 mlog_errno(ret);
1389                 goto out;
1390         }
1391
1392         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), bh,
1393                                       OCFS2_JOURNAL_ACCESS_WRITE);
1394         if (ret < 0) {
1395                 mlog_errno(ret);
1396                 goto out_trans;
1397         }
1398
1399         inode->i_mode &= ~S_ISUID;
1400         if ((inode->i_mode & S_ISGID) && (inode->i_mode & S_IXGRP))
1401                 inode->i_mode &= ~S_ISGID;
1402
1403         di = (struct ocfs2_dinode *) bh->b_data;
1404         di->i_mode = cpu_to_le16(inode->i_mode);
1405         ocfs2_update_inode_fsync_trans(handle, inode, 0);
1406
1407         ocfs2_journal_dirty(handle, bh);
1408
1409 out_trans:
1410         ocfs2_commit_trans(osb, handle);
1411 out:
1412         return ret;
1413 }
1414
1415 static int ocfs2_write_remove_suid(struct inode *inode)
1416 {
1417         int ret;
1418         struct buffer_head *bh = NULL;
1419
1420         ret = ocfs2_read_inode_block(inode, &bh);
1421         if (ret < 0) {
1422                 mlog_errno(ret);
1423                 goto out;
1424         }
1425
1426         ret =  __ocfs2_write_remove_suid(inode, bh);
1427 out:
1428         brelse(bh);
1429         return ret;
1430 }
1431
1432 /*
1433  * Allocate enough extents to cover the region starting at byte offset
1434  * start for len bytes. Existing extents are skipped, any extents
1435  * added are marked as "unwritten".
1436  */
1437 static int ocfs2_allocate_unwritten_extents(struct inode *inode,
1438                                             u64 start, u64 len)
1439 {
1440         int ret;
1441         u32 cpos, phys_cpos, clusters, alloc_size;
1442         u64 end = start + len;
1443         struct buffer_head *di_bh = NULL;
1444
1445         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1446                 ret = ocfs2_read_inode_block(inode, &di_bh);
1447                 if (ret) {
1448                         mlog_errno(ret);
1449                         goto out;
1450                 }
1451
1452                 /*
1453                  * Nothing to do if the requested reservation range
1454                  * fits within the inode.
1455                  */
1456                 if (ocfs2_size_fits_inline_data(di_bh, end))
1457                         goto out;
1458
1459                 ret = ocfs2_convert_inline_data_to_extents(inode, di_bh);
1460                 if (ret) {
1461                         mlog_errno(ret);
1462                         goto out;
1463                 }
1464         }
1465
1466         /*
1467          * We consider both start and len to be inclusive.
1468          */
1469         cpos = start >> OCFS2_SB(inode->i_sb)->s_clustersize_bits;
1470         clusters = ocfs2_clusters_for_bytes(inode->i_sb, start + len);
1471         clusters -= cpos;
1472
1473         while (clusters) {
1474                 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos,
1475                                          &alloc_size, NULL);
1476                 if (ret) {
1477                         mlog_errno(ret);
1478                         goto out;
1479                 }
1480
1481                 /*
1482                  * Hole or existing extent len can be arbitrary, so
1483                  * cap it to our own allocation request.
1484                  */
1485                 if (alloc_size > clusters)
1486                         alloc_size = clusters;
1487
1488                 if (phys_cpos) {
1489                         /*
1490                          * We already have an allocation at this
1491                          * region so we can safely skip it.
1492                          */
1493                         goto next;
1494                 }
1495
1496                 ret = __ocfs2_extend_allocation(inode, cpos, alloc_size, 1);
1497                 if (ret) {
1498                         if (ret != -ENOSPC)
1499                                 mlog_errno(ret);
1500                         goto out;
1501                 }
1502
1503 next:
1504                 cpos += alloc_size;
1505                 clusters -= alloc_size;
1506         }
1507
1508         ret = 0;
1509 out:
1510
1511         brelse(di_bh);
1512         return ret;
1513 }
1514
1515 /*
1516  * Truncate a byte range, avoiding pages within partial clusters. This
1517  * preserves those pages for the zeroing code to write to.
1518  */
1519 static void ocfs2_truncate_cluster_pages(struct inode *inode, u64 byte_start,
1520                                          u64 byte_len)
1521 {
1522         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1523         loff_t start, end;
1524         struct address_space *mapping = inode->i_mapping;
1525
1526         start = (loff_t)ocfs2_align_bytes_to_clusters(inode->i_sb, byte_start);
1527         end = byte_start + byte_len;
1528         end = end & ~(osb->s_clustersize - 1);
1529
1530         if (start < end) {
1531                 unmap_mapping_range(mapping, start, end - start, 0);
1532                 truncate_inode_pages_range(mapping, start, end - 1);
1533         }
1534 }
1535
1536 static int ocfs2_zero_partial_clusters(struct inode *inode,
1537                                        u64 start, u64 len)
1538 {
1539         int ret = 0;
1540         u64 tmpend = 0;
1541         u64 end = start + len;
1542         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1543         unsigned int csize = osb->s_clustersize;
1544         handle_t *handle;
1545
1546         /*
1547          * The "start" and "end" values are NOT necessarily part of
1548          * the range whose allocation is being deleted. Rather, this
1549          * is what the user passed in with the request. We must zero
1550          * partial clusters here. There's no need to worry about
1551          * physical allocation - the zeroing code knows to skip holes.
1552          */
1553         trace_ocfs2_zero_partial_clusters(
1554                 (unsigned long long)OCFS2_I(inode)->ip_blkno,
1555                 (unsigned long long)start, (unsigned long long)end);
1556
1557         /*
1558          * If both edges are on a cluster boundary then there's no
1559          * zeroing required as the region is part of the allocation to
1560          * be truncated.
1561          */
1562         if ((start & (csize - 1)) == 0 && (end & (csize - 1)) == 0)
1563                 goto out;
1564
1565         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1566         if (IS_ERR(handle)) {
1567                 ret = PTR_ERR(handle);
1568                 mlog_errno(ret);
1569                 goto out;
1570         }
1571
1572         /*
1573          * If start is on a cluster boundary and end is somewhere in another
1574          * cluster, we have not COWed the cluster starting at start, unless
1575          * end is also within the same cluster. So, in this case, we skip this
1576          * first call to ocfs2_zero_range_for_truncate() truncate and move on
1577          * to the next one.
1578          */
1579         if ((start & (csize - 1)) != 0) {
1580                 /*
1581                  * We want to get the byte offset of the end of the 1st
1582                  * cluster.
1583                  */
1584                 tmpend = (u64)osb->s_clustersize +
1585                         (start & ~(osb->s_clustersize - 1));
1586                 if (tmpend > end)
1587                         tmpend = end;
1588
1589                 trace_ocfs2_zero_partial_clusters_range1(
1590                         (unsigned long long)start,
1591                         (unsigned long long)tmpend);
1592
1593                 ret = ocfs2_zero_range_for_truncate(inode, handle, start,
1594                                                     tmpend);
1595                 if (ret)
1596                         mlog_errno(ret);
1597         }
1598
1599         if (tmpend < end) {
1600                 /*
1601                  * This may make start and end equal, but the zeroing
1602                  * code will skip any work in that case so there's no
1603                  * need to catch it up here.
1604                  */
1605                 start = end & ~(osb->s_clustersize - 1);
1606
1607                 trace_ocfs2_zero_partial_clusters_range2(
1608                         (unsigned long long)start, (unsigned long long)end);
1609
1610                 ret = ocfs2_zero_range_for_truncate(inode, handle, start, end);
1611                 if (ret)
1612                         mlog_errno(ret);
1613         }
1614         ocfs2_update_inode_fsync_trans(handle, inode, 1);
1615
1616         ocfs2_commit_trans(osb, handle);
1617 out:
1618         return ret;
1619 }
1620
1621 static int ocfs2_find_rec(struct ocfs2_extent_list *el, u32 pos)
1622 {
1623         int i;
1624         struct ocfs2_extent_rec *rec = NULL;
1625
1626         for (i = le16_to_cpu(el->l_next_free_rec) - 1; i >= 0; i--) {
1627
1628                 rec = &el->l_recs[i];
1629
1630                 if (le32_to_cpu(rec->e_cpos) < pos)
1631                         break;
1632         }
1633
1634         return i;
1635 }
1636
1637 /*
1638  * Helper to calculate the punching pos and length in one run, we handle the
1639  * following three cases in order:
1640  *
1641  * - remove the entire record
1642  * - remove a partial record
1643  * - no record needs to be removed (hole-punching completed)
1644 */
1645 static void ocfs2_calc_trunc_pos(struct inode *inode,
1646                                  struct ocfs2_extent_list *el,
1647                                  struct ocfs2_extent_rec *rec,
1648                                  u32 trunc_start, u32 *trunc_cpos,
1649                                  u32 *trunc_len, u32 *trunc_end,
1650                                  u64 *blkno, int *done)
1651 {
1652         int ret = 0;
1653         u32 coff, range;
1654
1655         range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
1656
1657         if (le32_to_cpu(rec->e_cpos) >= trunc_start) {
1658                 /*
1659                  * remove an entire extent record.
1660                  */
1661                 *trunc_cpos = le32_to_cpu(rec->e_cpos);
1662                 /*
1663                  * Skip holes if any.
1664                  */
1665                 if (range < *trunc_end)
1666                         *trunc_end = range;
1667                 *trunc_len = *trunc_end - le32_to_cpu(rec->e_cpos);
1668                 *blkno = le64_to_cpu(rec->e_blkno);
1669                 *trunc_end = le32_to_cpu(rec->e_cpos);
1670         } else if (range > trunc_start) {
1671                 /*
1672                  * remove a partial extent record, which means we're
1673                  * removing the last extent record.
1674                  */
1675                 *trunc_cpos = trunc_start;
1676                 /*
1677                  * skip hole if any.
1678                  */
1679                 if (range < *trunc_end)
1680                         *trunc_end = range;
1681                 *trunc_len = *trunc_end - trunc_start;
1682                 coff = trunc_start - le32_to_cpu(rec->e_cpos);
1683                 *blkno = le64_to_cpu(rec->e_blkno) +
1684                                 ocfs2_clusters_to_blocks(inode->i_sb, coff);
1685                 *trunc_end = trunc_start;
1686         } else {
1687                 /*
1688                  * It may have two following possibilities:
1689                  *
1690                  * - last record has been removed
1691                  * - trunc_start was within a hole
1692                  *
1693                  * both two cases mean the completion of hole punching.
1694                  */
1695                 ret = 1;
1696         }
1697
1698         *done = ret;
1699 }
1700
1701 int ocfs2_remove_inode_range(struct inode *inode,
1702                              struct buffer_head *di_bh, u64 byte_start,
1703                              u64 byte_len)
1704 {
1705         int ret = 0, flags = 0, done = 0, i;
1706         u32 trunc_start, trunc_len, trunc_end, trunc_cpos, phys_cpos;
1707         u32 cluster_in_el;
1708         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1709         struct ocfs2_cached_dealloc_ctxt dealloc;
1710         struct address_space *mapping = inode->i_mapping;
1711         struct ocfs2_extent_tree et;
1712         struct ocfs2_path *path = NULL;
1713         struct ocfs2_extent_list *el = NULL;
1714         struct ocfs2_extent_rec *rec = NULL;
1715         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
1716         u64 blkno, refcount_loc = le64_to_cpu(di->i_refcount_loc);
1717
1718         ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), di_bh);
1719         ocfs2_init_dealloc_ctxt(&dealloc);
1720
1721         trace_ocfs2_remove_inode_range(
1722                         (unsigned long long)OCFS2_I(inode)->ip_blkno,
1723                         (unsigned long long)byte_start,
1724                         (unsigned long long)byte_len);
1725
1726         if (byte_len == 0)
1727                 return 0;
1728
1729         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1730                 ret = ocfs2_truncate_inline(inode, di_bh, byte_start,
1731                                             byte_start + byte_len, 0);
1732                 if (ret) {
1733                         mlog_errno(ret);
1734                         goto out;
1735                 }
1736                 /*
1737                  * There's no need to get fancy with the page cache
1738                  * truncate of an inline-data inode. We're talking
1739                  * about less than a page here, which will be cached
1740                  * in the dinode buffer anyway.
1741                  */
1742                 unmap_mapping_range(mapping, 0, 0, 0);
1743                 truncate_inode_pages(mapping, 0);
1744                 goto out;
1745         }
1746
1747         /*
1748          * For reflinks, we may need to CoW 2 clusters which might be
1749          * partially zero'd later, if hole's start and end offset were
1750          * within one cluster(means is not exactly aligned to clustersize).
1751          */
1752
1753         if (ocfs2_is_refcount_inode(inode)) {
1754                 ret = ocfs2_cow_file_pos(inode, di_bh, byte_start);
1755                 if (ret) {
1756                         mlog_errno(ret);
1757                         goto out;
1758                 }
1759
1760                 ret = ocfs2_cow_file_pos(inode, di_bh, byte_start + byte_len);
1761                 if (ret) {
1762                         mlog_errno(ret);
1763                         goto out;
1764                 }
1765         }
1766
1767         trunc_start = ocfs2_clusters_for_bytes(osb->sb, byte_start);
1768         trunc_end = (byte_start + byte_len) >> osb->s_clustersize_bits;
1769         cluster_in_el = trunc_end;
1770
1771         ret = ocfs2_zero_partial_clusters(inode, byte_start, byte_len);
1772         if (ret) {
1773                 mlog_errno(ret);
1774                 goto out;
1775         }
1776
1777         path = ocfs2_new_path_from_et(&et);
1778         if (!path) {
1779                 ret = -ENOMEM;
1780                 mlog_errno(ret);
1781                 goto out;
1782         }
1783
1784         while (trunc_end > trunc_start) {
1785
1786                 ret = ocfs2_find_path(INODE_CACHE(inode), path,
1787                                       cluster_in_el);
1788                 if (ret) {
1789                         mlog_errno(ret);
1790                         goto out;
1791                 }
1792
1793                 el = path_leaf_el(path);
1794
1795                 i = ocfs2_find_rec(el, trunc_end);
1796                 /*
1797                  * Need to go to previous extent block.
1798                  */
1799                 if (i < 0) {
1800                         if (path->p_tree_depth == 0)
1801                                 break;
1802
1803                         ret = ocfs2_find_cpos_for_left_leaf(inode->i_sb,
1804                                                             path,
1805                                                             &cluster_in_el);
1806                         if (ret) {
1807                                 mlog_errno(ret);
1808                                 goto out;
1809                         }
1810
1811                         /*
1812                          * We've reached the leftmost extent block,
1813                          * it's safe to leave.
1814                          */
1815                         if (cluster_in_el == 0)
1816                                 break;
1817
1818                         /*
1819                          * The 'pos' searched for previous extent block is
1820                          * always one cluster less than actual trunc_end.
1821                          */
1822                         trunc_end = cluster_in_el + 1;
1823
1824                         ocfs2_reinit_path(path, 1);
1825
1826                         continue;
1827
1828                 } else
1829                         rec = &el->l_recs[i];
1830
1831                 ocfs2_calc_trunc_pos(inode, el, rec, trunc_start, &trunc_cpos,
1832                                      &trunc_len, &trunc_end, &blkno, &done);
1833                 if (done)
1834                         break;
1835
1836                 flags = rec->e_flags;
1837                 phys_cpos = ocfs2_blocks_to_clusters(inode->i_sb, blkno);
1838
1839                 ret = ocfs2_remove_btree_range(inode, &et, trunc_cpos,
1840                                                phys_cpos, trunc_len, flags,
1841                                                &dealloc, refcount_loc, false);
1842                 if (ret < 0) {
1843                         mlog_errno(ret);
1844                         goto out;
1845                 }
1846
1847                 cluster_in_el = trunc_end;
1848
1849                 ocfs2_reinit_path(path, 1);
1850         }
1851
1852         ocfs2_truncate_cluster_pages(inode, byte_start, byte_len);
1853
1854 out:
1855         ocfs2_free_path(path);
1856         ocfs2_schedule_truncate_log_flush(osb, 1);
1857         ocfs2_run_deallocs(osb, &dealloc);
1858
1859         return ret;
1860 }
1861
1862 /*
1863  * Parts of this function taken from xfs_change_file_space()
1864  */
1865 static int __ocfs2_change_file_space(struct file *file, struct inode *inode,
1866                                      loff_t f_pos, unsigned int cmd,
1867                                      struct ocfs2_space_resv *sr,
1868                                      int change_size)
1869 {
1870         int ret;
1871         s64 llen;
1872         loff_t size;
1873         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1874         struct buffer_head *di_bh = NULL;
1875         handle_t *handle;
1876         unsigned long long max_off = inode->i_sb->s_maxbytes;
1877
1878         if (ocfs2_is_hard_readonly(osb) || ocfs2_is_soft_readonly(osb))
1879                 return -EROFS;
1880
1881         inode_lock(inode);
1882
1883         /*
1884          * This prevents concurrent writes on other nodes
1885          */
1886         ret = ocfs2_rw_lock(inode, 1);
1887         if (ret) {
1888                 mlog_errno(ret);
1889                 goto out;
1890         }
1891
1892         ret = ocfs2_inode_lock(inode, &di_bh, 1);
1893         if (ret) {
1894                 mlog_errno(ret);
1895                 goto out_rw_unlock;
1896         }
1897
1898         if (inode->i_flags & (S_IMMUTABLE|S_APPEND)) {
1899                 ret = -EPERM;
1900                 goto out_inode_unlock;
1901         }
1902
1903         switch (sr->l_whence) {
1904         case 0: /*SEEK_SET*/
1905                 break;
1906         case 1: /*SEEK_CUR*/
1907                 sr->l_start += f_pos;
1908                 break;
1909         case 2: /*SEEK_END*/
1910                 sr->l_start += i_size_read(inode);
1911                 break;
1912         default:
1913                 ret = -EINVAL;
1914                 goto out_inode_unlock;
1915         }
1916         sr->l_whence = 0;
1917
1918         llen = sr->l_len > 0 ? sr->l_len - 1 : sr->l_len;
1919
1920         if (sr->l_start < 0
1921             || sr->l_start > max_off
1922             || (sr->l_start + llen) < 0
1923             || (sr->l_start + llen) > max_off) {
1924                 ret = -EINVAL;
1925                 goto out_inode_unlock;
1926         }
1927         size = sr->l_start + sr->l_len;
1928
1929         if (cmd == OCFS2_IOC_RESVSP || cmd == OCFS2_IOC_RESVSP64 ||
1930             cmd == OCFS2_IOC_UNRESVSP || cmd == OCFS2_IOC_UNRESVSP64) {
1931                 if (sr->l_len <= 0) {
1932                         ret = -EINVAL;
1933                         goto out_inode_unlock;
1934                 }
1935         }
1936
1937         if (file && should_remove_suid(file->f_path.dentry)) {
1938                 ret = __ocfs2_write_remove_suid(inode, di_bh);
1939                 if (ret) {
1940                         mlog_errno(ret);
1941                         goto out_inode_unlock;
1942                 }
1943         }
1944
1945         down_write(&OCFS2_I(inode)->ip_alloc_sem);
1946         switch (cmd) {
1947         case OCFS2_IOC_RESVSP:
1948         case OCFS2_IOC_RESVSP64:
1949                 /*
1950                  * This takes unsigned offsets, but the signed ones we
1951                  * pass have been checked against overflow above.
1952                  */
1953                 ret = ocfs2_allocate_unwritten_extents(inode, sr->l_start,
1954                                                        sr->l_len);
1955                 break;
1956         case OCFS2_IOC_UNRESVSP:
1957         case OCFS2_IOC_UNRESVSP64:
1958                 ret = ocfs2_remove_inode_range(inode, di_bh, sr->l_start,
1959                                                sr->l_len);
1960                 break;
1961         default:
1962                 ret = -EINVAL;
1963         }
1964         up_write(&OCFS2_I(inode)->ip_alloc_sem);
1965         if (ret) {
1966                 mlog_errno(ret);
1967                 goto out_inode_unlock;
1968         }
1969
1970         /*
1971          * We update c/mtime for these changes
1972          */
1973         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1974         if (IS_ERR(handle)) {
1975                 ret = PTR_ERR(handle);
1976                 mlog_errno(ret);
1977                 goto out_inode_unlock;
1978         }
1979
1980         if (change_size && i_size_read(inode) < size)
1981                 i_size_write(inode, size);
1982
1983         inode->i_ctime = inode->i_mtime = current_time(inode);
1984         ret = ocfs2_mark_inode_dirty(handle, inode, di_bh);
1985         if (ret < 0)
1986                 mlog_errno(ret);
1987
1988         if (file && (file->f_flags & O_SYNC))
1989                 handle->h_sync = 1;
1990
1991         ocfs2_commit_trans(osb, handle);
1992
1993 out_inode_unlock:
1994         brelse(di_bh);
1995         ocfs2_inode_unlock(inode, 1);
1996 out_rw_unlock:
1997         ocfs2_rw_unlock(inode, 1);
1998
1999 out:
2000         inode_unlock(inode);
2001         return ret;
2002 }
2003
2004 int ocfs2_change_file_space(struct file *file, unsigned int cmd,
2005                             struct ocfs2_space_resv *sr)
2006 {
2007         struct inode *inode = file_inode(file);
2008         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
2009         int ret;
2010
2011         if ((cmd == OCFS2_IOC_RESVSP || cmd == OCFS2_IOC_RESVSP64) &&
2012             !ocfs2_writes_unwritten_extents(osb))
2013                 return -ENOTTY;
2014         else if ((cmd == OCFS2_IOC_UNRESVSP || cmd == OCFS2_IOC_UNRESVSP64) &&
2015                  !ocfs2_sparse_alloc(osb))
2016                 return -ENOTTY;
2017
2018         if (!S_ISREG(inode->i_mode))
2019                 return -EINVAL;
2020
2021         if (!(file->f_mode & FMODE_WRITE))
2022                 return -EBADF;
2023
2024         ret = mnt_want_write_file(file);
2025         if (ret)
2026                 return ret;
2027         ret = __ocfs2_change_file_space(file, inode, file->f_pos, cmd, sr, 0);
2028         mnt_drop_write_file(file);
2029         return ret;
2030 }
2031
2032 static long ocfs2_fallocate(struct file *file, int mode, loff_t offset,
2033                             loff_t len)
2034 {
2035         struct inode *inode = file_inode(file);
2036         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
2037         struct ocfs2_space_resv sr;
2038         int change_size = 1;
2039         int cmd = OCFS2_IOC_RESVSP64;
2040
2041         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
2042                 return -EOPNOTSUPP;
2043         if (!ocfs2_writes_unwritten_extents(osb))
2044                 return -EOPNOTSUPP;
2045
2046         if (mode & FALLOC_FL_KEEP_SIZE)
2047                 change_size = 0;
2048
2049         if (mode & FALLOC_FL_PUNCH_HOLE)
2050                 cmd = OCFS2_IOC_UNRESVSP64;
2051
2052         sr.l_whence = 0;
2053         sr.l_start = (s64)offset;
2054         sr.l_len = (s64)len;
2055
2056         return __ocfs2_change_file_space(NULL, inode, offset, cmd, &sr,
2057                                          change_size);
2058 }
2059
2060 int ocfs2_check_range_for_refcount(struct inode *inode, loff_t pos,
2061                                    size_t count)
2062 {
2063         int ret = 0;
2064         unsigned int extent_flags;
2065         u32 cpos, clusters, extent_len, phys_cpos;
2066         struct super_block *sb = inode->i_sb;
2067
2068         if (!ocfs2_refcount_tree(OCFS2_SB(inode->i_sb)) ||
2069             !ocfs2_is_refcount_inode(inode) ||
2070             OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
2071                 return 0;
2072
2073         cpos = pos >> OCFS2_SB(sb)->s_clustersize_bits;
2074         clusters = ocfs2_clusters_for_bytes(sb, pos + count) - cpos;
2075
2076         while (clusters) {
2077                 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos, &extent_len,
2078                                          &extent_flags);
2079                 if (ret < 0) {
2080                         mlog_errno(ret);
2081                         goto out;
2082                 }
2083
2084                 if (phys_cpos && (extent_flags & OCFS2_EXT_REFCOUNTED)) {
2085                         ret = 1;
2086                         break;
2087                 }
2088
2089                 if (extent_len > clusters)
2090                         extent_len = clusters;
2091
2092                 clusters -= extent_len;
2093                 cpos += extent_len;
2094         }
2095 out:
2096         return ret;
2097 }
2098
2099 static int ocfs2_is_io_unaligned(struct inode *inode, size_t count, loff_t pos)
2100 {
2101         int blockmask = inode->i_sb->s_blocksize - 1;
2102         loff_t final_size = pos + count;
2103
2104         if ((pos & blockmask) || (final_size & blockmask))
2105                 return 1;
2106         return 0;
2107 }
2108
2109 static int ocfs2_prepare_inode_for_refcount(struct inode *inode,
2110                                             struct file *file,
2111                                             loff_t pos, size_t count,
2112                                             int *meta_level)
2113 {
2114         int ret;
2115         struct buffer_head *di_bh = NULL;
2116         u32 cpos = pos >> OCFS2_SB(inode->i_sb)->s_clustersize_bits;
2117         u32 clusters =
2118                 ocfs2_clusters_for_bytes(inode->i_sb, pos + count) - cpos;
2119
2120         ret = ocfs2_inode_lock(inode, &di_bh, 1);
2121         if (ret) {
2122                 mlog_errno(ret);
2123                 goto out;
2124         }
2125
2126         *meta_level = 1;
2127
2128         ret = ocfs2_refcount_cow(inode, di_bh, cpos, clusters, UINT_MAX);
2129         if (ret)
2130                 mlog_errno(ret);
2131 out:
2132         brelse(di_bh);
2133         return ret;
2134 }
2135
2136 static int ocfs2_prepare_inode_for_write(struct file *file,
2137                                          loff_t pos,
2138                                          size_t count)
2139 {
2140         int ret = 0, meta_level = 0;
2141         struct dentry *dentry = file->f_path.dentry;
2142         struct inode *inode = d_inode(dentry);
2143         loff_t end;
2144
2145         /*
2146          * We start with a read level meta lock and only jump to an ex
2147          * if we need to make modifications here.
2148          */
2149         for(;;) {
2150                 ret = ocfs2_inode_lock(inode, NULL, meta_level);
2151                 if (ret < 0) {
2152                         meta_level = -1;
2153                         mlog_errno(ret);
2154                         goto out;
2155                 }
2156
2157                 /* Clear suid / sgid if necessary. We do this here
2158                  * instead of later in the write path because
2159                  * remove_suid() calls ->setattr without any hint that
2160                  * we may have already done our cluster locking. Since
2161                  * ocfs2_setattr() *must* take cluster locks to
2162                  * proceed, this will lead us to recursively lock the
2163                  * inode. There's also the dinode i_size state which
2164                  * can be lost via setattr during extending writes (we
2165                  * set inode->i_size at the end of a write. */
2166                 if (should_remove_suid(dentry)) {
2167                         if (meta_level == 0) {
2168                                 ocfs2_inode_unlock(inode, meta_level);
2169                                 meta_level = 1;
2170                                 continue;
2171                         }
2172
2173                         ret = ocfs2_write_remove_suid(inode);
2174                         if (ret < 0) {
2175                                 mlog_errno(ret);
2176                                 goto out_unlock;
2177                         }
2178                 }
2179
2180                 end = pos + count;
2181
2182                 ret = ocfs2_check_range_for_refcount(inode, pos, count);
2183                 if (ret == 1) {
2184                         ocfs2_inode_unlock(inode, meta_level);
2185                         meta_level = -1;
2186
2187                         ret = ocfs2_prepare_inode_for_refcount(inode,
2188                                                                file,
2189                                                                pos,
2190                                                                count,
2191                                                                &meta_level);
2192                 }
2193
2194                 if (ret < 0) {
2195                         mlog_errno(ret);
2196                         goto out_unlock;
2197                 }
2198
2199                 break;
2200         }
2201
2202 out_unlock:
2203         trace_ocfs2_prepare_inode_for_write(OCFS2_I(inode)->ip_blkno,
2204                                             pos, count);
2205
2206         if (meta_level >= 0)
2207                 ocfs2_inode_unlock(inode, meta_level);
2208
2209 out:
2210         return ret;
2211 }
2212
2213 static ssize_t ocfs2_file_write_iter(struct kiocb *iocb,
2214                                     struct iov_iter *from)
2215 {
2216         int direct_io, rw_level;
2217         ssize_t written = 0;
2218         ssize_t ret;
2219         size_t count = iov_iter_count(from);
2220         struct file *file = iocb->ki_filp;
2221         struct inode *inode = file_inode(file);
2222         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
2223         int full_coherency = !(osb->s_mount_opt &
2224                                OCFS2_MOUNT_COHERENCY_BUFFERED);
2225         void *saved_ki_complete = NULL;
2226         int append_write = ((iocb->ki_pos + count) >=
2227                         i_size_read(inode) ? 1 : 0);
2228
2229         trace_ocfs2_file_aio_write(inode, file, file->f_path.dentry,
2230                 (unsigned long long)OCFS2_I(inode)->ip_blkno,
2231                 file->f_path.dentry->d_name.len,
2232                 file->f_path.dentry->d_name.name,
2233                 (unsigned int)from->nr_segs);   /* GRRRRR */
2234
2235         if (count == 0)
2236                 return 0;
2237
2238         direct_io = iocb->ki_flags & IOCB_DIRECT ? 1 : 0;
2239
2240         inode_lock(inode);
2241
2242         /*
2243          * Concurrent O_DIRECT writes are allowed with
2244          * mount_option "coherency=buffered".
2245          * For append write, we must take rw EX.
2246          */
2247         rw_level = (!direct_io || full_coherency || append_write);
2248
2249         ret = ocfs2_rw_lock(inode, rw_level);
2250         if (ret < 0) {
2251                 mlog_errno(ret);
2252                 goto out_mutex;
2253         }
2254
2255         /*
2256          * O_DIRECT writes with "coherency=full" need to take EX cluster
2257          * inode_lock to guarantee coherency.
2258          */
2259         if (direct_io && full_coherency) {
2260                 /*
2261                  * We need to take and drop the inode lock to force
2262                  * other nodes to drop their caches.  Buffered I/O
2263                  * already does this in write_begin().
2264                  */
2265                 ret = ocfs2_inode_lock(inode, NULL, 1);
2266                 if (ret < 0) {
2267                         mlog_errno(ret);
2268                         goto out;
2269                 }
2270
2271                 ocfs2_inode_unlock(inode, 1);
2272         }
2273
2274         ret = generic_write_checks(iocb, from);
2275         if (ret <= 0) {
2276                 if (ret)
2277                         mlog_errno(ret);
2278                 goto out;
2279         }
2280         count = ret;
2281
2282         ret = ocfs2_prepare_inode_for_write(file, iocb->ki_pos, count);
2283         if (ret < 0) {
2284                 mlog_errno(ret);
2285                 goto out;
2286         }
2287
2288         if (direct_io && !is_sync_kiocb(iocb) &&
2289             ocfs2_is_io_unaligned(inode, count, iocb->ki_pos)) {
2290                 /*
2291                  * Make it a sync io if it's an unaligned aio.
2292                  */
2293                 saved_ki_complete = xchg(&iocb->ki_complete, NULL);
2294         }
2295
2296         /* communicate with ocfs2_dio_end_io */
2297         ocfs2_iocb_set_rw_locked(iocb, rw_level);
2298
2299         written = __generic_file_write_iter(iocb, from);
2300         /* buffered aio wouldn't have proper lock coverage today */
2301         BUG_ON(written == -EIOCBQUEUED && !(iocb->ki_flags & IOCB_DIRECT));
2302
2303         /*
2304          * deep in g_f_a_w_n()->ocfs2_direct_IO we pass in a ocfs2_dio_end_io
2305          * function pointer which is called when o_direct io completes so that
2306          * it can unlock our rw lock.
2307          * Unfortunately there are error cases which call end_io and others
2308          * that don't.  so we don't have to unlock the rw_lock if either an
2309          * async dio is going to do it in the future or an end_io after an
2310          * error has already done it.
2311          */
2312         if ((written == -EIOCBQUEUED) || (!ocfs2_iocb_is_rw_locked(iocb))) {
2313                 rw_level = -1;
2314         }
2315
2316         if (unlikely(written <= 0))
2317                 goto out;
2318
2319         if (((file->f_flags & O_DSYNC) && !direct_io) ||
2320             IS_SYNC(inode)) {
2321                 ret = filemap_fdatawrite_range(file->f_mapping,
2322                                                iocb->ki_pos - written,
2323                                                iocb->ki_pos - 1);
2324                 if (ret < 0)
2325                         written = ret;
2326
2327                 if (!ret) {
2328                         ret = jbd2_journal_force_commit(osb->journal->j_journal);
2329                         if (ret < 0)
2330                                 written = ret;
2331                 }
2332
2333                 if (!ret)
2334                         ret = filemap_fdatawait_range(file->f_mapping,
2335                                                       iocb->ki_pos - written,
2336                                                       iocb->ki_pos - 1);
2337         }
2338
2339 out:
2340         if (saved_ki_complete)
2341                 xchg(&iocb->ki_complete, saved_ki_complete);
2342
2343         if (rw_level != -1)
2344                 ocfs2_rw_unlock(inode, rw_level);
2345
2346 out_mutex:
2347         inode_unlock(inode);
2348
2349         if (written)
2350                 ret = written;
2351         return ret;
2352 }
2353
2354 static ssize_t ocfs2_file_read_iter(struct kiocb *iocb,
2355                                    struct iov_iter *to)
2356 {
2357         int ret = 0, rw_level = -1, lock_level = 0;
2358         struct file *filp = iocb->ki_filp;
2359         struct inode *inode = file_inode(filp);
2360
2361         trace_ocfs2_file_aio_read(inode, filp, filp->f_path.dentry,
2362                         (unsigned long long)OCFS2_I(inode)->ip_blkno,
2363                         filp->f_path.dentry->d_name.len,
2364                         filp->f_path.dentry->d_name.name,
2365                         to->nr_segs);   /* GRRRRR */
2366
2367
2368         if (!inode) {
2369                 ret = -EINVAL;
2370                 mlog_errno(ret);
2371                 goto bail;
2372         }
2373
2374         /*
2375          * buffered reads protect themselves in ->readpage().  O_DIRECT reads
2376          * need locks to protect pending reads from racing with truncate.
2377          */
2378         if (iocb->ki_flags & IOCB_DIRECT) {
2379                 ret = ocfs2_rw_lock(inode, 0);
2380                 if (ret < 0) {
2381                         mlog_errno(ret);
2382                         goto bail;
2383                 }
2384                 rw_level = 0;
2385                 /* communicate with ocfs2_dio_end_io */
2386                 ocfs2_iocb_set_rw_locked(iocb, rw_level);
2387         }
2388
2389         /*
2390          * We're fine letting folks race truncates and extending
2391          * writes with read across the cluster, just like they can
2392          * locally. Hence no rw_lock during read.
2393          *
2394          * Take and drop the meta data lock to update inode fields
2395          * like i_size. This allows the checks down below
2396          * generic_file_aio_read() a chance of actually working.
2397          */
2398         ret = ocfs2_inode_lock_atime(inode, filp->f_path.mnt, &lock_level);
2399         if (ret < 0) {
2400                 mlog_errno(ret);
2401                 goto bail;
2402         }
2403         ocfs2_inode_unlock(inode, lock_level);
2404
2405         ret = generic_file_read_iter(iocb, to);
2406         trace_generic_file_aio_read_ret(ret);
2407
2408         /* buffered aio wouldn't have proper lock coverage today */
2409         BUG_ON(ret == -EIOCBQUEUED && !(iocb->ki_flags & IOCB_DIRECT));
2410
2411         /* see ocfs2_file_write_iter */
2412         if (ret == -EIOCBQUEUED || !ocfs2_iocb_is_rw_locked(iocb)) {
2413                 rw_level = -1;
2414         }
2415
2416 bail:
2417         if (rw_level != -1)
2418                 ocfs2_rw_unlock(inode, rw_level);
2419
2420         return ret;
2421 }
2422
2423 /* Refer generic_file_llseek_unlocked() */
2424 static loff_t ocfs2_file_llseek(struct file *file, loff_t offset, int whence)
2425 {
2426         struct inode *inode = file->f_mapping->host;
2427         int ret = 0;
2428
2429         inode_lock(inode);
2430
2431         switch (whence) {
2432         case SEEK_SET:
2433                 break;
2434         case SEEK_END:
2435                 /* SEEK_END requires the OCFS2 inode lock for the file
2436                  * because it references the file's size.
2437                  */
2438                 ret = ocfs2_inode_lock(inode, NULL, 0);
2439                 if (ret < 0) {
2440                         mlog_errno(ret);
2441                         goto out;
2442                 }
2443                 offset += i_size_read(inode);
2444                 ocfs2_inode_unlock(inode, 0);
2445                 break;
2446         case SEEK_CUR:
2447                 if (offset == 0) {
2448                         offset = file->f_pos;
2449                         goto out;
2450                 }
2451                 offset += file->f_pos;
2452                 break;
2453         case SEEK_DATA:
2454         case SEEK_HOLE:
2455                 ret = ocfs2_seek_data_hole_offset(file, &offset, whence);
2456                 if (ret)
2457                         goto out;
2458                 break;
2459         default:
2460                 ret = -EINVAL;
2461                 goto out;
2462         }
2463
2464         offset = vfs_setpos(file, offset, inode->i_sb->s_maxbytes);
2465
2466 out:
2467         inode_unlock(inode);
2468         if (ret)
2469                 return ret;
2470         return offset;
2471 }
2472
2473 static int ocfs2_file_clone_range(struct file *file_in,
2474                                   loff_t pos_in,
2475                                   struct file *file_out,
2476                                   loff_t pos_out,
2477                                   u64 len)
2478 {
2479         return ocfs2_reflink_remap_range(file_in, pos_in, file_out, pos_out,
2480                                          len, false);
2481 }
2482
2483 static ssize_t ocfs2_file_dedupe_range(struct file *src_file,
2484                                        u64 loff,
2485                                        u64 len,
2486                                        struct file *dst_file,
2487                                        u64 dst_loff)
2488 {
2489         int error;
2490
2491         error = ocfs2_reflink_remap_range(src_file, loff, dst_file, dst_loff,
2492                                           len, true);
2493         if (error)
2494                 return error;
2495         return len;
2496 }
2497
2498 const struct inode_operations ocfs2_file_iops = {
2499         .setattr        = ocfs2_setattr,
2500         .getattr        = ocfs2_getattr,
2501         .permission     = ocfs2_permission,
2502         .listxattr      = ocfs2_listxattr,
2503         .fiemap         = ocfs2_fiemap,
2504         .get_acl        = ocfs2_iop_get_acl,
2505         .set_acl        = ocfs2_iop_set_acl,
2506 };
2507
2508 const struct inode_operations ocfs2_special_file_iops = {
2509         .setattr        = ocfs2_setattr,
2510         .getattr        = ocfs2_getattr,
2511         .permission     = ocfs2_permission,
2512         .get_acl        = ocfs2_iop_get_acl,
2513         .set_acl        = ocfs2_iop_set_acl,
2514 };
2515
2516 /*
2517  * Other than ->lock, keep ocfs2_fops and ocfs2_dops in sync with
2518  * ocfs2_fops_no_plocks and ocfs2_dops_no_plocks!
2519  */
2520 const struct file_operations ocfs2_fops = {
2521         .llseek         = ocfs2_file_llseek,
2522         .mmap           = ocfs2_mmap,
2523         .fsync          = ocfs2_sync_file,
2524         .release        = ocfs2_file_release,
2525         .open           = ocfs2_file_open,
2526         .read_iter      = ocfs2_file_read_iter,
2527         .write_iter     = ocfs2_file_write_iter,
2528         .unlocked_ioctl = ocfs2_ioctl,
2529 #ifdef CONFIG_COMPAT
2530         .compat_ioctl   = ocfs2_compat_ioctl,
2531 #endif
2532         .lock           = ocfs2_lock,
2533         .flock          = ocfs2_flock,
2534         .splice_read    = generic_file_splice_read,
2535         .splice_write   = iter_file_splice_write,
2536         .fallocate      = ocfs2_fallocate,
2537         .clone_file_range = ocfs2_file_clone_range,
2538         .dedupe_file_range = ocfs2_file_dedupe_range,
2539 };
2540
2541 const struct file_operations ocfs2_dops = {
2542         .llseek         = generic_file_llseek,
2543         .read           = generic_read_dir,
2544         .iterate        = ocfs2_readdir,
2545         .fsync          = ocfs2_sync_file,
2546         .release        = ocfs2_dir_release,
2547         .open           = ocfs2_dir_open,
2548         .unlocked_ioctl = ocfs2_ioctl,
2549 #ifdef CONFIG_COMPAT
2550         .compat_ioctl   = ocfs2_compat_ioctl,
2551 #endif
2552         .lock           = ocfs2_lock,
2553         .flock          = ocfs2_flock,
2554 };
2555
2556 /*
2557  * POSIX-lockless variants of our file_operations.
2558  *
2559  * These will be used if the underlying cluster stack does not support
2560  * posix file locking, if the user passes the "localflocks" mount
2561  * option, or if we have a local-only fs.
2562  *
2563  * ocfs2_flock is in here because all stacks handle UNIX file locks,
2564  * so we still want it in the case of no stack support for
2565  * plocks. Internally, it will do the right thing when asked to ignore
2566  * the cluster.
2567  */
2568 const struct file_operations ocfs2_fops_no_plocks = {
2569         .llseek         = ocfs2_file_llseek,
2570         .mmap           = ocfs2_mmap,
2571         .fsync          = ocfs2_sync_file,
2572         .release        = ocfs2_file_release,
2573         .open           = ocfs2_file_open,
2574         .read_iter      = ocfs2_file_read_iter,
2575         .write_iter     = ocfs2_file_write_iter,
2576         .unlocked_ioctl = ocfs2_ioctl,
2577 #ifdef CONFIG_COMPAT
2578         .compat_ioctl   = ocfs2_compat_ioctl,
2579 #endif
2580         .flock          = ocfs2_flock,
2581         .splice_read    = generic_file_splice_read,
2582         .splice_write   = iter_file_splice_write,
2583         .fallocate      = ocfs2_fallocate,
2584         .clone_file_range = ocfs2_file_clone_range,
2585         .dedupe_file_range = ocfs2_file_dedupe_range,
2586 };
2587
2588 const struct file_operations ocfs2_dops_no_plocks = {
2589         .llseek         = generic_file_llseek,
2590         .read           = generic_read_dir,
2591         .iterate        = ocfs2_readdir,
2592         .fsync          = ocfs2_sync_file,
2593         .release        = ocfs2_dir_release,
2594         .open           = ocfs2_dir_open,
2595         .unlocked_ioctl = ocfs2_ioctl,
2596 #ifdef CONFIG_COMPAT
2597         .compat_ioctl   = ocfs2_compat_ioctl,
2598 #endif
2599         .flock          = ocfs2_flock,
2600 };