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 = file_write_and_wait_range(file, 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 & SB_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 /*
 717  * While a write will already be ordering the data, a truncate will not.
 718  * Thus, we need to explicitly order the zeroed pages.
 719  */
 720 static handle_t *ocfs2_zero_start_ordered_transaction(struct inode *inode,
 721                                                 struct buffer_head *di_bh)
 722 {
 723         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 724         handle_t *handle = NULL;
 725         int ret = 0;
 726
 727         if (!ocfs2_should_order_data(inode))
 728                 goto out;
 729
 730         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
 731         if (IS_ERR(handle)) {
 732                 ret = -ENOMEM;
 733                 mlog_errno(ret);
 734                 goto out;
 735         }
 736
 737         ret = ocfs2_jbd2_file_inode(handle, inode);
 738         if (ret < 0) {
 739                 mlog_errno(ret);
 740                 goto out;
 741         }
 742
 743         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
 744                                       OCFS2_JOURNAL_ACCESS_WRITE);
 745         if (ret)
 746                 mlog_errno(ret);
 747         ocfs2_update_inode_fsync_trans(handle, inode, 1);
 748
 749 out:
 750         if (ret) {
 751                 if (!IS_ERR(handle))
 752                         ocfs2_commit_trans(osb, handle);
 753                 handle = ERR_PTR(ret);
 754         }
 755         return handle;
 756 }
 757
 758 /* Some parts of this taken from generic_cont_expand, which turned out
 759  * to be too fragile to do exactly what we need without us having to
 760  * worry about recursive locking in ->write_begin() and ->write_end(). */
 761 static int ocfs2_write_zero_page(struct inode *inode, u64 abs_from,
 762                                  u64 abs_to, struct buffer_head *di_bh)
 763 {
 764         struct address_space *mapping = inode->i_mapping;
 765         struct page *page;
 766         unsigned long index = abs_from >> PAGE_SHIFT;
 767         handle_t *handle;
 768         int ret = 0;
 769         unsigned zero_from, zero_to, block_start, block_end;
 770         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
 771
 772         BUG_ON(abs_from >= abs_to);
 773         BUG_ON(abs_to > (((u64)index + 1) << PAGE_SHIFT));
 774         BUG_ON(abs_from & (inode->i_blkbits - 1));
 775
 776         handle = ocfs2_zero_start_ordered_transaction(inode, di_bh);
 777         if (IS_ERR(handle)) {
 778                 ret = PTR_ERR(handle);
 779                 goto out;
 780         }
 781
 782         page = find_or_create_page(mapping, index, GFP_NOFS);
 783         if (!page) {
 784                 ret = -ENOMEM;
 785                 mlog_errno(ret);
 786                 goto out_commit_trans;
 787         }
 788
 789         /* Get the offsets within the page that we want to zero */
 790         zero_from = abs_from & (PAGE_SIZE - 1);
 791         zero_to = abs_to & (PAGE_SIZE - 1);
 792         if (!zero_to)
 793                 zero_to = PAGE_SIZE;
 794
 795         trace_ocfs2_write_zero_page(
 796                         (unsigned long long)OCFS2_I(inode)->ip_blkno,
 797                         (unsigned long long)abs_from,
 798                         (unsigned long long)abs_to,
 799                         index, zero_from, zero_to);
 800
 801         /* We know that zero_from is block aligned */
 802         for (block_start = zero_from; block_start < zero_to;
 803              block_start = block_end) {
 804                 block_end = block_start + i_blocksize(inode);
 805
 806                 /*
 807                  * block_start is block-aligned.  Bump it by one to force
 808                  * __block_write_begin and block_commit_write to zero the
 809                  * whole block.
 810                  */
 811                 ret = __block_write_begin(page, block_start + 1, 0,
 812                                           ocfs2_get_block);
 813                 if (ret < 0) {
 814                         mlog_errno(ret);
 815                         goto out_unlock;
 816                 }
 817
 818
 819                 /* must not update i_size! */
 820                 ret = block_commit_write(page, block_start + 1,
 821                                          block_start + 1);
 822                 if (ret < 0)
 823                         mlog_errno(ret);
 824                 else
 825                         ret = 0;
 826         }
 827
 828         /*
 829          * fs-writeback will release the dirty pages without page lock
 830          * whose offset are over inode size, the release happens at
 831          * block_write_full_page().
 832          */
 833         i_size_write(inode, abs_to);
 834         inode->i_blocks = ocfs2_inode_sector_count(inode);
 835         di->i_size = cpu_to_le64((u64)i_size_read(inode));
 836         inode->i_mtime = inode->i_ctime = current_time(inode);
 837         di->i_mtime = di->i_ctime = cpu_to_le64(inode->i_mtime.tv_sec);
 838         di->i_ctime_nsec = cpu_to_le32(inode->i_mtime.tv_nsec);
 839         di->i_mtime_nsec = di->i_ctime_nsec;
 840         if (handle) {
 841                 ocfs2_journal_dirty(handle, di_bh);
 842                 ocfs2_update_inode_fsync_trans(handle, inode, 1);
 843         }
 844
 845 out_unlock:
 846         unlock_page(page);
 847         put_page(page);
 848 out_commit_trans:
 849         if (handle)
 850                 ocfs2_commit_trans(OCFS2_SB(inode->i_sb), handle);
 851 out:
 852         return ret;
 853 }
 854
 855 /*
 856  * Find the next range to zero.  We do this in terms of bytes because
 857  * that's what ocfs2_zero_extend() wants, and it is dealing with the
 858  * pagecache.  We may return multiple extents.
 859  *
 860  * zero_start and zero_end are ocfs2_zero_extend()s current idea of what
 861  * needs to be zeroed.  range_start and range_end return the next zeroing
 862  * range.  A subsequent call should pass the previous range_end as its
 863  * zero_start.  If range_end is 0, there's nothing to do.
 864  *
 865  * Unwritten extents are skipped over.  Refcounted extents are CoWd.
 866  */
 867 static int ocfs2_zero_extend_get_range(struct inode *inode,
 868                                        struct buffer_head *di_bh,
 869                                        u64 zero_start, u64 zero_end,
 870                                        u64 *range_start, u64 *range_end)
 871 {
 872         int rc = 0, needs_cow = 0;
 873         u32 p_cpos, zero_clusters = 0;
 874         u32 zero_cpos =
 875                 zero_start >> OCFS2_SB(inode->i_sb)->s_clustersize_bits;
 876         u32 last_cpos = ocfs2_clusters_for_bytes(inode->i_sb, zero_end);
 877         unsigned int num_clusters = 0;
 878         unsigned int ext_flags = 0;
 879
 880         while (zero_cpos < last_cpos) {
 881                 rc = ocfs2_get_clusters(inode, zero_cpos, &p_cpos,
 882                                         &num_clusters, &ext_flags);
 883                 if (rc) {
 884                         mlog_errno(rc);
 885                         goto out;
 886                 }
 887
 888                 if (p_cpos && !(ext_flags & OCFS2_EXT_UNWRITTEN)) {
 889                         zero_clusters = num_clusters;
 890                         if (ext_flags & OCFS2_EXT_REFCOUNTED)
 891                                 needs_cow = 1;
 892                         break;
 893                 }
 894
 895                 zero_cpos += num_clusters;
 896         }
 897         if (!zero_clusters) {
 898                 *range_end = 0;
 899                 goto out;
 900         }
 901
 902         while ((zero_cpos + zero_clusters) < last_cpos) {
 903                 rc = ocfs2_get_clusters(inode, zero_cpos + zero_clusters,
 904                                         &p_cpos, &num_clusters,
 905                                         &ext_flags);
 906                 if (rc) {
 907                         mlog_errno(rc);
 908                         goto out;
 909                 }
 910
 911                 if (!p_cpos || (ext_flags & OCFS2_EXT_UNWRITTEN))
 912                         break;
 913                 if (ext_flags & OCFS2_EXT_REFCOUNTED)
 914                         needs_cow = 1;
 915                 zero_clusters += num_clusters;
 916         }
 917         if ((zero_cpos + zero_clusters) > last_cpos)
 918                 zero_clusters = last_cpos - zero_cpos;
 919
 920         if (needs_cow) {
 921                 rc = ocfs2_refcount_cow(inode, di_bh, zero_cpos,
 922                                         zero_clusters, UINT_MAX);
 923                 if (rc) {
 924                         mlog_errno(rc);
 925                         goto out;
 926                 }
 927         }
 928
 929         *range_start = ocfs2_clusters_to_bytes(inode->i_sb, zero_cpos);
 930         *range_end = ocfs2_clusters_to_bytes(inode->i_sb,
 931                                              zero_cpos + zero_clusters);
 932
 933 out:
 934         return rc;
 935 }
 936
 937 /*
 938  * Zero one range returned from ocfs2_zero_extend_get_range().  The caller
 939  * has made sure that the entire range needs zeroing.
 940  */
 941 static int ocfs2_zero_extend_range(struct inode *inode, u64 range_start,
 942                                    u64 range_end, struct buffer_head *di_bh)
 943 {
 944         int rc = 0;
 945         u64 next_pos;
 946         u64 zero_pos = range_start;
 947
 948         trace_ocfs2_zero_extend_range(
 949                         (unsigned long long)OCFS2_I(inode)->ip_blkno,
 950                         (unsigned long long)range_start,
 951                         (unsigned long long)range_end);
 952         BUG_ON(range_start >= range_end);
 953
 954         while (zero_pos < range_end) {
 955                 next_pos = (zero_pos & PAGE_MASK) + PAGE_SIZE;
 956                 if (next_pos > range_end)
 957                         next_pos = range_end;
 958                 rc = ocfs2_write_zero_page(inode, zero_pos, next_pos, di_bh);
 959                 if (rc < 0) {
 960                         mlog_errno(rc);
 961                         break;
 962                 }
 963                 zero_pos = next_pos;
 964
 965                 /*
 966                  * Very large extends have the potential to lock up
 967                  * the cpu for extended periods of time.
 968                  */
 969                 cond_resched();
 970         }
 971
 972         return rc;
 973 }
 974
 975 int ocfs2_zero_extend(struct inode *inode, struct buffer_head *di_bh,
 976                       loff_t zero_to_size)
 977 {
 978         int ret = 0;
 979         u64 zero_start, range_start = 0, range_end = 0;
 980         struct super_block *sb = inode->i_sb;
 981
 982         zero_start = ocfs2_align_bytes_to_blocks(sb, i_size_read(inode));
 983         trace_ocfs2_zero_extend((unsigned long long)OCFS2_I(inode)->ip_blkno,
 984                                 (unsigned long long)zero_start,
 985                                 (unsigned long long)i_size_read(inode));
 986         while (zero_start < zero_to_size) {
 987                 ret = ocfs2_zero_extend_get_range(inode, di_bh, zero_start,
 988                                                   zero_to_size,
 989                                                   &range_start,
 990                                                   &range_end);
 991                 if (ret) {
 992                         mlog_errno(ret);
 993                         break;
 994                 }
 995                 if (!range_end)
 996                         break;
 997                 /* Trim the ends */
 998                 if (range_start < zero_start)
 999                         range_start = zero_start;
1000                 if (range_end > zero_to_size)
1001                         range_end = zero_to_size;
1002
1003                 ret = ocfs2_zero_extend_range(inode, range_start,
1004                                               range_end, di_bh);
1005                 if (ret) {
1006                         mlog_errno(ret);
1007                         break;
1008                 }
1009                 zero_start = range_end;
1010         }
1011
1012         return ret;
1013 }
1014
1015 int ocfs2_extend_no_holes(struct inode *inode, struct buffer_head *di_bh,
1016                           u64 new_i_size, u64 zero_to)
1017 {
1018         int ret;
1019         u32 clusters_to_add;
1020         struct ocfs2_inode_info *oi = OCFS2_I(inode);
1021
1022         /*
1023          * Only quota files call this without a bh, and they can't be
1024          * refcounted.
1025          */
1026         BUG_ON(!di_bh && ocfs2_is_refcount_inode(inode));
1027         BUG_ON(!di_bh && !(oi->ip_flags & OCFS2_INODE_SYSTEM_FILE));
1028
1029         clusters_to_add = ocfs2_clusters_for_bytes(inode->i_sb, new_i_size);
1030         if (clusters_to_add < oi->ip_clusters)
1031                 clusters_to_add = 0;
1032         else
1033                 clusters_to_add -= oi->ip_clusters;
1034
1035         if (clusters_to_add) {
1036                 ret = __ocfs2_extend_allocation(inode, oi->ip_clusters,
1037                                                 clusters_to_add, 0);
1038                 if (ret) {
1039                         mlog_errno(ret);
1040                         goto out;
1041                 }
1042         }
1043
1044         /*
1045          * Call this even if we don't add any clusters to the tree. We
1046          * still need to zero the area between the old i_size and the
1047          * new i_size.
1048          */
1049         ret = ocfs2_zero_extend(inode, di_bh, zero_to);
1050         if (ret < 0)
1051                 mlog_errno(ret);
1052
1053 out:
1054         return ret;
1055 }
1056
1057 static int ocfs2_extend_file(struct inode *inode,
1058                              struct buffer_head *di_bh,
1059                              u64 new_i_size)
1060 {
1061         int ret = 0;
1062         struct ocfs2_inode_info *oi = OCFS2_I(inode);
1063
1064         BUG_ON(!di_bh);
1065
1066         /* setattr sometimes calls us like this. */
1067         if (new_i_size == 0)
1068                 goto out;
1069
1070         if (i_size_read(inode) == new_i_size)
1071                 goto out;
1072         BUG_ON(new_i_size < i_size_read(inode));
1073
1074         /*
1075          * The alloc sem blocks people in read/write from reading our
1076          * allocation until we're done changing it. We depend on
1077          * i_mutex to block other extend/truncate calls while we're
1078          * here.  We even have to hold it for sparse files because there
1079          * might be some tail zeroing.
1080          */
1081         down_write(&oi->ip_alloc_sem);
1082
1083         if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1084                 /*
1085                  * We can optimize small extends by keeping the inodes
1086                  * inline data.
1087                  */
1088                 if (ocfs2_size_fits_inline_data(di_bh, new_i_size)) {
1089                         up_write(&oi->ip_alloc_sem);
1090                         goto out_update_size;
1091                 }
1092
1093                 ret = ocfs2_convert_inline_data_to_extents(inode, di_bh);
1094                 if (ret) {
1095                         up_write(&oi->ip_alloc_sem);
1096                         mlog_errno(ret);
1097                         goto out;
1098                 }
1099         }
1100
1101         if (ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)))
1102                 ret = ocfs2_zero_extend(inode, di_bh, new_i_size);
1103         else
1104                 ret = ocfs2_extend_no_holes(inode, di_bh, new_i_size,
1105                                             new_i_size);
1106
1107         up_write(&oi->ip_alloc_sem);
1108
1109         if (ret < 0) {
1110                 mlog_errno(ret);
1111                 goto out;
1112         }
1113
1114 out_update_size:
1115         ret = ocfs2_simple_size_update(inode, di_bh, new_i_size);
1116         if (ret < 0)
1117                 mlog_errno(ret);
1118
1119 out:
1120         return ret;
1121 }
1122
1123 int ocfs2_setattr(struct dentry *dentry, struct iattr *attr)
1124 {
1125         int status = 0, size_change;
1126         int inode_locked = 0;
1127         struct inode *inode = d_inode(dentry);
1128         struct super_block *sb = inode->i_sb;
1129         struct ocfs2_super *osb = OCFS2_SB(sb);
1130         struct buffer_head *bh = NULL;
1131         handle_t *handle = NULL;
1132         struct dquot *transfer_to[MAXQUOTAS] = { };
1133         int qtype;
1134         int had_lock;
1135         struct ocfs2_lock_holder oh;
1136
1137         trace_ocfs2_setattr(inode, dentry,
1138                             (unsigned long long)OCFS2_I(inode)->ip_blkno,
1139                             dentry->d_name.len, dentry->d_name.name,
1140                             attr->ia_valid, attr->ia_mode,
1141                             from_kuid(&init_user_ns, attr->ia_uid),
1142                             from_kgid(&init_user_ns, attr->ia_gid));
1143
1144         /* ensuring we don't even attempt to truncate a symlink */
1145         if (S_ISLNK(inode->i_mode))
1146                 attr->ia_valid &= ~ATTR_SIZE;
1147
1148 #define OCFS2_VALID_ATTRS (ATTR_ATIME | ATTR_MTIME | ATTR_CTIME | ATTR_SIZE \
1149                            | ATTR_GID | ATTR_UID | ATTR_MODE)
1150         if (!(attr->ia_valid & OCFS2_VALID_ATTRS))
1151                 return 0;
1152
1153         status = setattr_prepare(dentry, attr);
1154         if (status)
1155                 return status;
1156
1157         if (is_quota_modification(inode, attr)) {
1158                 status = dquot_initialize(inode);
1159                 if (status)
1160                         return status;
1161         }
1162         size_change = S_ISREG(inode->i_mode) && attr->ia_valid & ATTR_SIZE;
1163         if (size_change) {
1164                 /*
1165                  * Here we should wait dio to finish before inode lock
1166                  * to avoid a deadlock between ocfs2_setattr() and
1167                  * ocfs2_dio_end_io_write()
1168                  */
1169                 inode_dio_wait(inode);
1170
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                 if (i_size_read(inode) >= attr->ia_size) {
1211                         if (ocfs2_should_order_data(inode)) {
1212                                 status = ocfs2_begin_ordered_truncate(inode,
1213                                                                       attr->ia_size);
1214                                 if (status)
1215                                         goto bail_unlock;
1216                         }
1217                         status = ocfs2_truncate_file(inode, bh, attr->ia_size);
1218                 } else
1219                         status = ocfs2_extend_file(inode, bh, attr->ia_size);
1220                 if (status < 0) {
1221                         if (status != -ENOSPC)
1222                                 mlog_errno(status);
1223                         status = -ENOSPC;
1224                         goto bail_unlock;
1225                 }
1226         }
1227
1228         if ((attr->ia_valid & ATTR_UID && !uid_eq(attr->ia_uid, inode->i_uid)) ||
1229             (attr->ia_valid & ATTR_GID && !gid_eq(attr->ia_gid, inode->i_gid))) {
1230                 /*
1231                  * Gather pointers to quota structures so that allocation /
1232                  * freeing of quota structures happens here and not inside
1233                  * dquot_transfer() where we have problems with lock ordering
1234                  */
1235                 if (attr->ia_valid & ATTR_UID && !uid_eq(attr->ia_uid, inode->i_uid)
1236                     && OCFS2_HAS_RO_COMPAT_FEATURE(sb,
1237                     OCFS2_FEATURE_RO_COMPAT_USRQUOTA)) {
1238                         transfer_to[USRQUOTA] = dqget(sb, make_kqid_uid(attr->ia_uid));
1239                         if (IS_ERR(transfer_to[USRQUOTA])) {
1240                                 status = PTR_ERR(transfer_to[USRQUOTA]);
1241                                 goto bail_unlock;
1242                         }
1243                 }
1244                 if (attr->ia_valid & ATTR_GID && !gid_eq(attr->ia_gid, inode->i_gid)
1245                     && OCFS2_HAS_RO_COMPAT_FEATURE(sb,
1246                     OCFS2_FEATURE_RO_COMPAT_GRPQUOTA)) {
1247                         transfer_to[GRPQUOTA] = dqget(sb, make_kqid_gid(attr->ia_gid));
1248                         if (IS_ERR(transfer_to[GRPQUOTA])) {
1249                                 status = PTR_ERR(transfer_to[GRPQUOTA]);
1250                                 goto bail_unlock;
1251                         }
1252                 }
1253                 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS +
1254                                            2 * ocfs2_quota_trans_credits(sb));
1255                 if (IS_ERR(handle)) {
1256                         status = PTR_ERR(handle);
1257                         mlog_errno(status);
1258                         goto bail_unlock;
1259                 }
1260                 status = __dquot_transfer(inode, transfer_to);
1261                 if (status < 0)
1262                         goto bail_commit;
1263         } else {
1264                 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1265                 if (IS_ERR(handle)) {
1266                         status = PTR_ERR(handle);
1267                         mlog_errno(status);
1268                         goto bail_unlock;
1269                 }
1270         }
1271
1272         setattr_copy(inode, attr);
1273         mark_inode_dirty(inode);
1274
1275         status = ocfs2_mark_inode_dirty(handle, inode, bh);
1276         if (status < 0)
1277                 mlog_errno(status);
1278
1279 bail_commit:
1280         ocfs2_commit_trans(osb, handle);
1281 bail_unlock:
1282         if (status && inode_locked) {
1283                 ocfs2_inode_unlock_tracker(inode, 1, &oh, had_lock);
1284                 inode_locked = 0;
1285         }
1286 bail_unlock_rw:
1287         if (size_change)
1288                 ocfs2_rw_unlock(inode, 1);
1289 bail:
1290
1291         /* Release quota pointers in case we acquired them */
1292         for (qtype = 0; qtype < OCFS2_MAXQUOTAS; qtype++)
1293                 dqput(transfer_to[qtype]);
1294
1295         if (!status && attr->ia_valid & ATTR_MODE) {
1296                 status = ocfs2_acl_chmod(inode, bh);
1297                 if (status < 0)
1298                         mlog_errno(status);
1299         }
1300         if (inode_locked)
1301                 ocfs2_inode_unlock_tracker(inode, 1, &oh, had_lock);
1302
1303         brelse(bh);
1304         return status;
1305 }
1306
1307 int ocfs2_getattr(const struct path *path, struct kstat *stat,
1308                   u32 request_mask, unsigned int flags)
1309 {
1310         struct inode *inode = d_inode(path->dentry);
1311         struct super_block *sb = path->dentry->d_sb;
1312         struct ocfs2_super *osb = sb->s_fs_info;
1313         int err;
1314
1315         err = ocfs2_inode_revalidate(path->dentry);
1316         if (err) {
1317                 if (err != -ENOENT)
1318                         mlog_errno(err);
1319                 goto bail;
1320         }
1321
1322         generic_fillattr(inode, stat);
1323         /*
1324          * If there is inline data in the inode, the inode will normally not
1325          * have data blocks allocated (it may have an external xattr block).
1326          * Report at least one sector for such files, so tools like tar, rsync,
1327          * others don't incorrectly think the file is completely sparse.
1328          */
1329         if (unlikely(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL))
1330                 stat->blocks += (stat->size + 511)>>9;
1331
1332         /* We set the blksize from the cluster size for performance */
1333         stat->blksize = osb->s_clustersize;
1334
1335 bail:
1336         return err;
1337 }
1338
1339 int ocfs2_permission(struct inode *inode, int mask)
1340 {
1341         int ret, had_lock;
1342         struct ocfs2_lock_holder oh;
1343
1344         if (mask & MAY_NOT_BLOCK)
1345                 return -ECHILD;
1346
1347         had_lock = ocfs2_inode_lock_tracker(inode, NULL, 0, &oh);
1348         if (had_lock < 0) {
1349                 ret = had_lock;
1350                 goto out;
1351         } else if (had_lock) {
1352                 /* See comments in ocfs2_setattr() for details.
1353                  * The call chain of this case could be:
1354                  * do_sys_open()
1355                  *  may_open()
1356                  *   inode_permission()
1357                  *    ocfs2_permission()
1358                  *     ocfs2_iop_get_acl()
1359                  */
1360                 mlog(ML_ERROR, "Another case of recursive locking:\n");
1361                 dump_stack();
1362         }
1363
1364         ret = generic_permission(inode, mask);
1365
1366         ocfs2_inode_unlock_tracker(inode, 0, &oh, had_lock);
1367 out:
1368         return ret;
1369 }
1370
1371 static int __ocfs2_write_remove_suid(struct inode *inode,
1372                                      struct buffer_head *bh)
1373 {
1374         int ret;
1375         handle_t *handle;
1376         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1377         struct ocfs2_dinode *di;
1378
1379         trace_ocfs2_write_remove_suid(
1380                         (unsigned long long)OCFS2_I(inode)->ip_blkno,
1381                         inode->i_mode);
1382
1383         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1384         if (IS_ERR(handle)) {
1385                 ret = PTR_ERR(handle);
1386                 mlog_errno(ret);
1387                 goto out;
1388         }
1389
1390         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), bh,
1391                                       OCFS2_JOURNAL_ACCESS_WRITE);
1392         if (ret < 0) {
1393                 mlog_errno(ret);
1394                 goto out_trans;
1395         }
1396
1397         inode->i_mode &= ~S_ISUID;
1398         if ((inode->i_mode & S_ISGID) && (inode->i_mode & S_IXGRP))
1399                 inode->i_mode &= ~S_ISGID;
1400
1401         di = (struct ocfs2_dinode *) bh->b_data;
1402         di->i_mode = cpu_to_le16(inode->i_mode);
1403         ocfs2_update_inode_fsync_trans(handle, inode, 0);
1404
1405         ocfs2_journal_dirty(handle, bh);
1406
1407 out_trans:
1408         ocfs2_commit_trans(osb, handle);
1409 out:
1410         return ret;
1411 }
1412
1413 static int ocfs2_write_remove_suid(struct inode *inode)
1414 {
1415         int ret;
1416         struct buffer_head *bh = NULL;
1417
1418         ret = ocfs2_read_inode_block(inode, &bh);
1419         if (ret < 0) {
1420                 mlog_errno(ret);
1421                 goto out;
1422         }
1423
1424         ret =  __ocfs2_write_remove_suid(inode, bh);
1425 out:
1426         brelse(bh);
1427         return ret;
1428 }
1429
1430 /*
1431  * Allocate enough extents to cover the region starting at byte offset
1432  * start for len bytes. Existing extents are skipped, any extents
1433  * added are marked as "unwritten".
1434  */
1435 static int ocfs2_allocate_unwritten_extents(struct inode *inode,
1436                                             u64 start, u64 len)
1437 {
1438         int ret;
1439         u32 cpos, phys_cpos, clusters, alloc_size;
1440         u64 end = start + len;
1441         struct buffer_head *di_bh = NULL;
1442
1443         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1444                 ret = ocfs2_read_inode_block(inode, &di_bh);
1445                 if (ret) {
1446                         mlog_errno(ret);
1447                         goto out;
1448                 }
1449
1450                 /*
1451                  * Nothing to do if the requested reservation range
1452                  * fits within the inode.
1453                  */
1454                 if (ocfs2_size_fits_inline_data(di_bh, end))
1455                         goto out;
1456
1457                 ret = ocfs2_convert_inline_data_to_extents(inode, di_bh);
1458                 if (ret) {
1459                         mlog_errno(ret);
1460                         goto out;
1461                 }
1462         }
1463
1464         /*
1465          * We consider both start and len to be inclusive.
1466          */
1467         cpos = start >> OCFS2_SB(inode->i_sb)->s_clustersize_bits;
1468         clusters = ocfs2_clusters_for_bytes(inode->i_sb, start + len);
1469         clusters -= cpos;
1470
1471         while (clusters) {
1472                 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos,
1473                                          &alloc_size, NULL);
1474                 if (ret) {
1475                         mlog_errno(ret);
1476                         goto out;
1477                 }
1478
1479                 /*
1480                  * Hole or existing extent len can be arbitrary, so
1481                  * cap it to our own allocation request.
1482                  */
1483                 if (alloc_size > clusters)
1484                         alloc_size = clusters;
1485
1486                 if (phys_cpos) {
1487                         /*
1488                          * We already have an allocation at this
1489                          * region so we can safely skip it.
1490                          */
1491                         goto next;
1492                 }
1493
1494                 ret = __ocfs2_extend_allocation(inode, cpos, alloc_size, 1);
1495                 if (ret) {
1496                         if (ret != -ENOSPC)
1497                                 mlog_errno(ret);
1498                         goto out;
1499                 }
1500
1501 next:
1502                 cpos += alloc_size;
1503                 clusters -= alloc_size;
1504         }
1505
1506         ret = 0;
1507 out:
1508
1509         brelse(di_bh);
1510         return ret;
1511 }
1512
1513 /*
1514  * Truncate a byte range, avoiding pages within partial clusters. This
1515  * preserves those pages for the zeroing code to write to.
1516  */
1517 static void ocfs2_truncate_cluster_pages(struct inode *inode, u64 byte_start,
1518                                          u64 byte_len)
1519 {
1520         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1521         loff_t start, end;
1522         struct address_space *mapping = inode->i_mapping;
1523
1524         start = (loff_t)ocfs2_align_bytes_to_clusters(inode->i_sb, byte_start);
1525         end = byte_start + byte_len;
1526         end = end & ~(osb->s_clustersize - 1);
1527
1528         if (start < end) {
1529                 unmap_mapping_range(mapping, start, end - start, 0);
1530                 truncate_inode_pages_range(mapping, start, end - 1);
1531         }
1532 }
1533
1534 static int ocfs2_zero_partial_clusters(struct inode *inode,
1535                                        u64 start, u64 len)
1536 {
1537         int ret = 0;
1538         u64 tmpend = 0;
1539         u64 end = start + len;
1540         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1541         unsigned int csize = osb->s_clustersize;
1542         handle_t *handle;
1543
1544         /*
1545          * The "start" and "end" values are NOT necessarily part of
1546          * the range whose allocation is being deleted. Rather, this
1547          * is what the user passed in with the request. We must zero
1548          * partial clusters here. There's no need to worry about
1549          * physical allocation - the zeroing code knows to skip holes.
1550          */
1551         trace_ocfs2_zero_partial_clusters(
1552                 (unsigned long long)OCFS2_I(inode)->ip_blkno,
1553                 (unsigned long long)start, (unsigned long long)end);
1554
1555         /*
1556          * If both edges are on a cluster boundary then there's no
1557          * zeroing required as the region is part of the allocation to
1558          * be truncated.
1559          */
1560         if ((start & (csize - 1)) == 0 && (end & (csize - 1)) == 0)
1561                 goto out;
1562
1563         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1564         if (IS_ERR(handle)) {
1565                 ret = PTR_ERR(handle);
1566                 mlog_errno(ret);
1567                 goto out;
1568         }
1569
1570         /*
1571          * If start is on a cluster boundary and end is somewhere in another
1572          * cluster, we have not COWed the cluster starting at start, unless
1573          * end is also within the same cluster. So, in this case, we skip this
1574          * first call to ocfs2_zero_range_for_truncate() truncate and move on
1575          * to the next one.
1576          */
1577         if ((start & (csize - 1)) != 0) {
1578                 /*
1579                  * We want to get the byte offset of the end of the 1st
1580                  * cluster.
1581                  */
1582                 tmpend = (u64)osb->s_clustersize +
1583                         (start & ~(osb->s_clustersize - 1));
1584                 if (tmpend > end)
1585                         tmpend = end;
1586
1587                 trace_ocfs2_zero_partial_clusters_range1(
1588                         (unsigned long long)start,
1589                         (unsigned long long)tmpend);
1590
1591                 ret = ocfs2_zero_range_for_truncate(inode, handle, start,
1592                                                     tmpend);
1593                 if (ret)
1594                         mlog_errno(ret);
1595         }
1596
1597         if (tmpend < end) {
1598                 /*
1599                  * This may make start and end equal, but the zeroing
1600                  * code will skip any work in that case so there's no
1601                  * need to catch it up here.
1602                  */
1603                 start = end & ~(osb->s_clustersize - 1);
1604
1605                 trace_ocfs2_zero_partial_clusters_range2(
1606                         (unsigned long long)start, (unsigned long long)end);
1607
1608                 ret = ocfs2_zero_range_for_truncate(inode, handle, start, end);
1609                 if (ret)
1610                         mlog_errno(ret);
1611         }
1612         ocfs2_update_inode_fsync_trans(handle, inode, 1);
1613
1614         ocfs2_commit_trans(osb, handle);
1615 out:
1616         return ret;
1617 }
1618
1619 static int ocfs2_find_rec(struct ocfs2_extent_list *el, u32 pos)
1620 {
1621         int i;
1622         struct ocfs2_extent_rec *rec = NULL;
1623
1624         for (i = le16_to_cpu(el->l_next_free_rec) - 1; i >= 0; i--) {
1625
1626                 rec = &el->l_recs[i];
1627
1628                 if (le32_to_cpu(rec->e_cpos) < pos)
1629                         break;
1630         }
1631
1632         return i;
1633 }
1634
1635 /*
1636  * Helper to calculate the punching pos and length in one run, we handle the
1637  * following three cases in order:
1638  *
1639  * - remove the entire record
1640  * - remove a partial record
1641  * - no record needs to be removed (hole-punching completed)
1642 */
1643 static void ocfs2_calc_trunc_pos(struct inode *inode,
1644                                  struct ocfs2_extent_list *el,
1645                                  struct ocfs2_extent_rec *rec,
1646                                  u32 trunc_start, u32 *trunc_cpos,
1647                                  u32 *trunc_len, u32 *trunc_end,
1648                                  u64 *blkno, int *done)
1649 {
1650         int ret = 0;
1651         u32 coff, range;
1652
1653         range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
1654
1655         if (le32_to_cpu(rec->e_cpos) >= trunc_start) {
1656                 /*
1657                  * remove an entire extent record.
1658                  */
1659                 *trunc_cpos = le32_to_cpu(rec->e_cpos);
1660                 /*
1661                  * Skip holes if any.
1662                  */
1663                 if (range < *trunc_end)
1664                         *trunc_end = range;
1665                 *trunc_len = *trunc_end - le32_to_cpu(rec->e_cpos);
1666                 *blkno = le64_to_cpu(rec->e_blkno);
1667                 *trunc_end = le32_to_cpu(rec->e_cpos);
1668         } else if (range > trunc_start) {
1669                 /*
1670                  * remove a partial extent record, which means we're
1671                  * removing the last extent record.
1672                  */
1673                 *trunc_cpos = trunc_start;
1674                 /*
1675                  * skip hole if any.
1676                  */
1677                 if (range < *trunc_end)
1678                         *trunc_end = range;
1679                 *trunc_len = *trunc_end - trunc_start;
1680                 coff = trunc_start - le32_to_cpu(rec->e_cpos);
1681                 *blkno = le64_to_cpu(rec->e_blkno) +
1682                                 ocfs2_clusters_to_blocks(inode->i_sb, coff);
1683                 *trunc_end = trunc_start;
1684         } else {
1685                 /*
1686                  * It may have two following possibilities:
1687                  *
1688                  * - last record has been removed
1689                  * - trunc_start was within a hole
1690                  *
1691                  * both two cases mean the completion of hole punching.
1692                  */
1693                 ret = 1;
1694         }
1695
1696         *done = ret;
1697 }
1698
1699 int ocfs2_remove_inode_range(struct inode *inode,
1700                              struct buffer_head *di_bh, u64 byte_start,
1701                              u64 byte_len)
1702 {
1703         int ret = 0, flags = 0, done = 0, i;
1704         u32 trunc_start, trunc_len, trunc_end, trunc_cpos, phys_cpos;
1705         u32 cluster_in_el;
1706         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1707         struct ocfs2_cached_dealloc_ctxt dealloc;
1708         struct address_space *mapping = inode->i_mapping;
1709         struct ocfs2_extent_tree et;
1710         struct ocfs2_path *path = NULL;
1711         struct ocfs2_extent_list *el = NULL;
1712         struct ocfs2_extent_rec *rec = NULL;
1713         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
1714         u64 blkno, refcount_loc = le64_to_cpu(di->i_refcount_loc);
1715
1716         ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), di_bh);
1717         ocfs2_init_dealloc_ctxt(&dealloc);
1718
1719         trace_ocfs2_remove_inode_range(
1720                         (unsigned long long)OCFS2_I(inode)->ip_blkno,
1721                         (unsigned long long)byte_start,
1722                         (unsigned long long)byte_len);
1723
1724         if (byte_len == 0)
1725                 return 0;
1726
1727         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1728                 ret = ocfs2_truncate_inline(inode, di_bh, byte_start,
1729                                             byte_start + byte_len, 0);
1730                 if (ret) {
1731                         mlog_errno(ret);
1732                         goto out;
1733                 }
1734                 /*
1735                  * There's no need to get fancy with the page cache
1736                  * truncate of an inline-data inode. We're talking
1737                  * about less than a page here, which will be cached
1738                  * in the dinode buffer anyway.
1739                  */
1740                 unmap_mapping_range(mapping, 0, 0, 0);
1741                 truncate_inode_pages(mapping, 0);
1742                 goto out;
1743         }
1744
1745         /*
1746          * For reflinks, we may need to CoW 2 clusters which might be
1747          * partially zero'd later, if hole's start and end offset were
1748          * within one cluster(means is not exactly aligned to clustersize).
1749          */
1750
1751         if (ocfs2_is_refcount_inode(inode)) {
1752                 ret = ocfs2_cow_file_pos(inode, di_bh, byte_start);
1753                 if (ret) {
1754                         mlog_errno(ret);
1755                         goto out;
1756                 }
1757
1758                 ret = ocfs2_cow_file_pos(inode, di_bh, byte_start + byte_len);
1759                 if (ret) {
1760                         mlog_errno(ret);
1761                         goto out;
1762                 }
1763         }
1764
1765         trunc_start = ocfs2_clusters_for_bytes(osb->sb, byte_start);
1766         trunc_end = (byte_start + byte_len) >> osb->s_clustersize_bits;
1767         cluster_in_el = trunc_end;
1768
1769         ret = ocfs2_zero_partial_clusters(inode, byte_start, byte_len);
1770         if (ret) {
1771                 mlog_errno(ret);
1772                 goto out;
1773         }
1774
1775         path = ocfs2_new_path_from_et(&et);
1776         if (!path) {
1777                 ret = -ENOMEM;
1778                 mlog_errno(ret);
1779                 goto out;
1780         }
1781
1782         while (trunc_end > trunc_start) {
1783
1784                 ret = ocfs2_find_path(INODE_CACHE(inode), path,
1785                                       cluster_in_el);
1786                 if (ret) {
1787                         mlog_errno(ret);
1788                         goto out;
1789                 }
1790
1791                 el = path_leaf_el(path);
1792
1793                 i = ocfs2_find_rec(el, trunc_end);
1794                 /*
1795                  * Need to go to previous extent block.
1796                  */
1797                 if (i < 0) {
1798                         if (path->p_tree_depth == 0)
1799                                 break;
1800
1801                         ret = ocfs2_find_cpos_for_left_leaf(inode->i_sb,
1802                                                             path,
1803                                                             &cluster_in_el);
1804                         if (ret) {
1805                                 mlog_errno(ret);
1806                                 goto out;
1807                         }
1808
1809                         /*
1810                          * We've reached the leftmost extent block,
1811                          * it's safe to leave.
1812                          */
1813                         if (cluster_in_el == 0)
1814                                 break;
1815
1816                         /*
1817                          * The 'pos' searched for previous extent block is
1818                          * always one cluster less than actual trunc_end.
1819                          */
1820                         trunc_end = cluster_in_el + 1;
1821
1822                         ocfs2_reinit_path(path, 1);
1823
1824                         continue;
1825
1826                 } else
1827                         rec = &el->l_recs[i];
1828
1829                 ocfs2_calc_trunc_pos(inode, el, rec, trunc_start, &trunc_cpos,
1830                                      &trunc_len, &trunc_end, &blkno, &done);
1831                 if (done)
1832                         break;
1833
1834                 flags = rec->e_flags;
1835                 phys_cpos = ocfs2_blocks_to_clusters(inode->i_sb, blkno);
1836
1837                 ret = ocfs2_remove_btree_range(inode, &et, trunc_cpos,
1838                                                phys_cpos, trunc_len, flags,
1839                                                &dealloc, refcount_loc, false);
1840                 if (ret < 0) {
1841                         mlog_errno(ret);
1842                         goto out;
1843                 }
1844
1845                 cluster_in_el = trunc_end;
1846
1847                 ocfs2_reinit_path(path, 1);
1848         }
1849
1850         ocfs2_truncate_cluster_pages(inode, byte_start, byte_len);
1851
1852 out:
1853         ocfs2_free_path(path);
1854         ocfs2_schedule_truncate_log_flush(osb, 1);
1855         ocfs2_run_deallocs(osb, &dealloc);
1856
1857         return ret;
1858 }
1859
1860 /*
1861  * Parts of this function taken from xfs_change_file_space()
1862  */
1863 static int __ocfs2_change_file_space(struct file *file, struct inode *inode,
1864                                      loff_t f_pos, unsigned int cmd,
1865                                      struct ocfs2_space_resv *sr,
1866                                      int change_size)
1867 {
1868         int ret;
1869         s64 llen;
1870         loff_t size;
1871         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1872         struct buffer_head *di_bh = NULL;
1873         handle_t *handle;
1874         unsigned long long max_off = inode->i_sb->s_maxbytes;
1875
1876         if (ocfs2_is_hard_readonly(osb) || ocfs2_is_soft_readonly(osb))
1877                 return -EROFS;
1878
1879         inode_lock(inode);
1880
1881         /*
1882          * This prevents concurrent writes on other nodes
1883          */
1884         ret = ocfs2_rw_lock(inode, 1);
1885         if (ret) {
1886                 mlog_errno(ret);
1887                 goto out;
1888         }
1889
1890         ret = ocfs2_inode_lock(inode, &di_bh, 1);
1891         if (ret) {
1892                 mlog_errno(ret);
1893                 goto out_rw_unlock;
1894         }
1895
1896         if (inode->i_flags & (S_IMMUTABLE|S_APPEND)) {
1897                 ret = -EPERM;
1898                 goto out_inode_unlock;
1899         }
1900
1901         switch (sr->l_whence) {
1902         case 0: /*SEEK_SET*/
1903                 break;
1904         case 1: /*SEEK_CUR*/
1905                 sr->l_start += f_pos;
1906                 break;
1907         case 2: /*SEEK_END*/
1908                 sr->l_start += i_size_read(inode);
1909                 break;
1910         default:
1911                 ret = -EINVAL;
1912                 goto out_inode_unlock;
1913         }
1914         sr->l_whence = 0;
1915
1916         llen = sr->l_len > 0 ? sr->l_len - 1 : sr->l_len;
1917
1918         if (sr->l_start < 0
1919             || sr->l_start > max_off
1920             || (sr->l_start + llen) < 0
1921             || (sr->l_start + llen) > max_off) {
1922                 ret = -EINVAL;
1923                 goto out_inode_unlock;
1924         }
1925         size = sr->l_start + sr->l_len;
1926
1927         if (cmd == OCFS2_IOC_RESVSP || cmd == OCFS2_IOC_RESVSP64 ||
1928             cmd == OCFS2_IOC_UNRESVSP || cmd == OCFS2_IOC_UNRESVSP64) {
1929                 if (sr->l_len <= 0) {
1930                         ret = -EINVAL;
1931                         goto out_inode_unlock;
1932                 }
1933         }
1934
1935         if (file && should_remove_suid(file->f_path.dentry)) {
1936                 ret = __ocfs2_write_remove_suid(inode, di_bh);
1937                 if (ret) {
1938                         mlog_errno(ret);
1939                         goto out_inode_unlock;
1940                 }
1941         }
1942
1943         down_write(&OCFS2_I(inode)->ip_alloc_sem);
1944         switch (cmd) {
1945         case OCFS2_IOC_RESVSP:
1946         case OCFS2_IOC_RESVSP64:
1947                 /*
1948                  * This takes unsigned offsets, but the signed ones we
1949                  * pass have been checked against overflow above.
1950                  */
1951                 ret = ocfs2_allocate_unwritten_extents(inode, sr->l_start,
1952                                                        sr->l_len);
1953                 break;
1954         case OCFS2_IOC_UNRESVSP:
1955         case OCFS2_IOC_UNRESVSP64:
1956                 ret = ocfs2_remove_inode_range(inode, di_bh, sr->l_start,
1957                                                sr->l_len);
1958                 break;
1959         default:
1960                 ret = -EINVAL;
1961         }
1962         up_write(&OCFS2_I(inode)->ip_alloc_sem);
1963         if (ret) {
1964                 mlog_errno(ret);
1965                 goto out_inode_unlock;
1966         }
1967
1968         /*
1969          * We update c/mtime for these changes
1970          */
1971         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1972         if (IS_ERR(handle)) {
1973                 ret = PTR_ERR(handle);
1974                 mlog_errno(ret);
1975                 goto out_inode_unlock;
1976         }
1977
1978         if (change_size && i_size_read(inode) < size)
1979                 i_size_write(inode, size);
1980
1981         inode->i_ctime = inode->i_mtime = current_time(inode);
1982         ret = ocfs2_mark_inode_dirty(handle, inode, di_bh);
1983         if (ret < 0)
1984                 mlog_errno(ret);
1985
1986         if (file && (file->f_flags & O_SYNC))
1987                 handle->h_sync = 1;
1988
1989         ocfs2_commit_trans(osb, handle);
1990
1991 out_inode_unlock:
1992         brelse(di_bh);
1993         ocfs2_inode_unlock(inode, 1);
1994 out_rw_unlock:
1995         ocfs2_rw_unlock(inode, 1);
1996
1997 out:
1998         inode_unlock(inode);
1999         return ret;
2000 }
2001
2002 int ocfs2_change_file_space(struct file *file, unsigned int cmd,
2003                             struct ocfs2_space_resv *sr)
2004 {
2005         struct inode *inode = file_inode(file);
2006         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
2007         int ret;
2008
2009         if ((cmd == OCFS2_IOC_RESVSP || cmd == OCFS2_IOC_RESVSP64) &&
2010             !ocfs2_writes_unwritten_extents(osb))
2011                 return -ENOTTY;
2012         else if ((cmd == OCFS2_IOC_UNRESVSP || cmd == OCFS2_IOC_UNRESVSP64) &&
2013                  !ocfs2_sparse_alloc(osb))
2014                 return -ENOTTY;
2015
2016         if (!S_ISREG(inode->i_mode))
2017                 return -EINVAL;
2018
2019         if (!(file->f_mode & FMODE_WRITE))
2020                 return -EBADF;
2021
2022         ret = mnt_want_write_file(file);
2023         if (ret)
2024                 return ret;
2025         ret = __ocfs2_change_file_space(file, inode, file->f_pos, cmd, sr, 0);
2026         mnt_drop_write_file(file);
2027         return ret;
2028 }
2029
2030 static long ocfs2_fallocate(struct file *file, int mode, loff_t offset,
2031                             loff_t len)
2032 {
2033         struct inode *inode = file_inode(file);
2034         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
2035         struct ocfs2_space_resv sr;
2036         int change_size = 1;
2037         int cmd = OCFS2_IOC_RESVSP64;
2038
2039         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
2040                 return -EOPNOTSUPP;
2041         if (!ocfs2_writes_unwritten_extents(osb))
2042                 return -EOPNOTSUPP;
2043
2044         if (mode & FALLOC_FL_KEEP_SIZE)
2045                 change_size = 0;
2046
2047         if (mode & FALLOC_FL_PUNCH_HOLE)
2048                 cmd = OCFS2_IOC_UNRESVSP64;
2049
2050         sr.l_whence = 0;
2051         sr.l_start = (s64)offset;
2052         sr.l_len = (s64)len;
2053
2054         return __ocfs2_change_file_space(NULL, inode, offset, cmd, &sr,
2055                                          change_size);
2056 }
2057
2058 int ocfs2_check_range_for_refcount(struct inode *inode, loff_t pos,
2059                                    size_t count)
2060 {
2061         int ret = 0;
2062         unsigned int extent_flags;
2063         u32 cpos, clusters, extent_len, phys_cpos;
2064         struct super_block *sb = inode->i_sb;
2065
2066         if (!ocfs2_refcount_tree(OCFS2_SB(inode->i_sb)) ||
2067             !ocfs2_is_refcount_inode(inode) ||
2068             OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
2069                 return 0;
2070
2071         cpos = pos >> OCFS2_SB(sb)->s_clustersize_bits;
2072         clusters = ocfs2_clusters_for_bytes(sb, pos + count) - cpos;
2073
2074         while (clusters) {
2075                 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos, &extent_len,
2076                                          &extent_flags);
2077                 if (ret < 0) {
2078                         mlog_errno(ret);
2079                         goto out;
2080                 }
2081
2082                 if (phys_cpos && (extent_flags & OCFS2_EXT_REFCOUNTED)) {
2083                         ret = 1;
2084                         break;
2085                 }
2086
2087                 if (extent_len > clusters)
2088                         extent_len = clusters;
2089
2090                 clusters -= extent_len;
2091                 cpos += extent_len;
2092         }
2093 out:
2094         return ret;
2095 }
2096
2097 static int ocfs2_is_io_unaligned(struct inode *inode, size_t count, loff_t pos)
2098 {
2099         int blockmask = inode->i_sb->s_blocksize - 1;
2100         loff_t final_size = pos + count;
2101
2102         if ((pos & blockmask) || (final_size & blockmask))
2103                 return 1;
2104         return 0;
2105 }
2106
2107 static int ocfs2_prepare_inode_for_refcount(struct inode *inode,
2108                                             struct file *file,
2109                                             loff_t pos, size_t count,
2110                                             int *meta_level)
2111 {
2112         int ret;
2113         struct buffer_head *di_bh = NULL;
2114         u32 cpos = pos >> OCFS2_SB(inode->i_sb)->s_clustersize_bits;
2115         u32 clusters =
2116                 ocfs2_clusters_for_bytes(inode->i_sb, pos + count) - cpos;
2117
2118         ret = ocfs2_inode_lock(inode, &di_bh, 1);
2119         if (ret) {
2120                 mlog_errno(ret);
2121                 goto out;
2122         }
2123
2124         *meta_level = 1;
2125
2126         ret = ocfs2_refcount_cow(inode, di_bh, cpos, clusters, UINT_MAX);
2127         if (ret)
2128                 mlog_errno(ret);
2129 out:
2130         brelse(di_bh);
2131         return ret;
2132 }
2133
2134 static int ocfs2_prepare_inode_for_write(struct file *file,
2135                                          loff_t pos,
2136                                          size_t count)
2137 {
2138         int ret = 0, meta_level = 0;
2139         struct dentry *dentry = file->f_path.dentry;
2140         struct inode *inode = d_inode(dentry);
2141         loff_t end;
2142
2143         /*
2144          * We start with a read level meta lock and only jump to an ex
2145          * if we need to make modifications here.
2146          */
2147         for(;;) {
2148                 ret = ocfs2_inode_lock(inode, NULL, meta_level);
2149                 if (ret < 0) {
2150                         meta_level = -1;
2151                         mlog_errno(ret);
2152                         goto out;
2153                 }
2154
2155                 /* Clear suid / sgid if necessary. We do this here
2156                  * instead of later in the write path because
2157                  * remove_suid() calls ->setattr without any hint that
2158                  * we may have already done our cluster locking. Since
2159                  * ocfs2_setattr() *must* take cluster locks to
2160                  * proceed, this will lead us to recursively lock the
2161                  * inode. There's also the dinode i_size state which
2162                  * can be lost via setattr during extending writes (we
2163                  * set inode->i_size at the end of a write. */
2164                 if (should_remove_suid(dentry)) {
2165                         if (meta_level == 0) {
2166                                 ocfs2_inode_unlock(inode, meta_level);
2167                                 meta_level = 1;
2168                                 continue;
2169                         }
2170
2171                         ret = ocfs2_write_remove_suid(inode);
2172                         if (ret < 0) {
2173                                 mlog_errno(ret);
2174                                 goto out_unlock;
2175                         }
2176                 }
2177
2178                 end = pos + count;
2179
2180                 ret = ocfs2_check_range_for_refcount(inode, pos, count);
2181                 if (ret == 1) {
2182                         ocfs2_inode_unlock(inode, meta_level);
2183                         meta_level = -1;
2184
2185                         ret = ocfs2_prepare_inode_for_refcount(inode,
2186                                                                file,
2187                                                                pos,
2188                                                                count,
2189                                                                &meta_level);
2190                 }
2191
2192                 if (ret < 0) {
2193                         mlog_errno(ret);
2194                         goto out_unlock;
2195                 }
2196
2197                 break;
2198         }
2199
2200 out_unlock:
2201         trace_ocfs2_prepare_inode_for_write(OCFS2_I(inode)->ip_blkno,
2202                                             pos, count);
2203
2204         if (meta_level >= 0)
2205                 ocfs2_inode_unlock(inode, meta_level);
2206
2207 out:
2208         return ret;
2209 }
2210
2211 static ssize_t ocfs2_file_write_iter(struct kiocb *iocb,
2212                                     struct iov_iter *from)
2213 {
2214         int direct_io, rw_level;
2215         ssize_t written = 0;
2216         ssize_t ret;
2217         size_t count = iov_iter_count(from);
2218         struct file *file = iocb->ki_filp;
2219         struct inode *inode = file_inode(file);
2220         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
2221         int full_coherency = !(osb->s_mount_opt &
2222                                OCFS2_MOUNT_COHERENCY_BUFFERED);
2223         void *saved_ki_complete = NULL;
2224         int append_write = ((iocb->ki_pos + count) >=
2225                         i_size_read(inode) ? 1 : 0);
2226
2227         trace_ocfs2_file_aio_write(inode, file, file->f_path.dentry,
2228                 (unsigned long long)OCFS2_I(inode)->ip_blkno,
2229                 file->f_path.dentry->d_name.len,
2230                 file->f_path.dentry->d_name.name,
2231                 (unsigned int)from->nr_segs);   /* GRRRRR */
2232
2233         if (count == 0)
2234                 return 0;
2235
2236         direct_io = iocb->ki_flags & IOCB_DIRECT ? 1 : 0;
2237
2238         inode_lock(inode);
2239
2240         /*
2241          * Concurrent O_DIRECT writes are allowed with
2242          * mount_option "coherency=buffered".
2243          * For append write, we must take rw EX.
2244          */
2245         rw_level = (!direct_io || full_coherency || append_write);
2246
2247         ret = ocfs2_rw_lock(inode, rw_level);
2248         if (ret < 0) {
2249                 mlog_errno(ret);
2250                 goto out_mutex;
2251         }
2252
2253         /*
2254          * O_DIRECT writes with "coherency=full" need to take EX cluster
2255          * inode_lock to guarantee coherency.
2256          */
2257         if (direct_io && full_coherency) {
2258                 /*
2259                  * We need to take and drop the inode lock to force
2260                  * other nodes to drop their caches.  Buffered I/O
2261                  * already does this in write_begin().
2262                  */
2263                 ret = ocfs2_inode_lock(inode, NULL, 1);
2264                 if (ret < 0) {
2265                         mlog_errno(ret);
2266                         goto out;
2267                 }
2268
2269                 ocfs2_inode_unlock(inode, 1);
2270         }
2271
2272         ret = generic_write_checks(iocb, from);
2273         if (ret <= 0) {
2274                 if (ret)
2275                         mlog_errno(ret);
2276                 goto out;
2277         }
2278         count = ret;
2279
2280         ret = ocfs2_prepare_inode_for_write(file, iocb->ki_pos, count);
2281         if (ret < 0) {
2282                 mlog_errno(ret);
2283                 goto out;
2284         }
2285
2286         if (direct_io && !is_sync_kiocb(iocb) &&
2287             ocfs2_is_io_unaligned(inode, count, iocb->ki_pos)) {
2288                 /*
2289                  * Make it a sync io if it's an unaligned aio.
2290                  */
2291                 saved_ki_complete = xchg(&iocb->ki_complete, NULL);
2292         }
2293
2294         /* communicate with ocfs2_dio_end_io */
2295         ocfs2_iocb_set_rw_locked(iocb, rw_level);
2296
2297         written = __generic_file_write_iter(iocb, from);
2298         /* buffered aio wouldn't have proper lock coverage today */
2299         BUG_ON(written == -EIOCBQUEUED && !(iocb->ki_flags & IOCB_DIRECT));
2300
2301         /*
2302          * deep in g_f_a_w_n()->ocfs2_direct_IO we pass in a ocfs2_dio_end_io
2303          * function pointer which is called when o_direct io completes so that
2304          * it can unlock our rw lock.
2305          * Unfortunately there are error cases which call end_io and others
2306          * that don't.  so we don't have to unlock the rw_lock if either an
2307          * async dio is going to do it in the future or an end_io after an
2308          * error has already done it.
2309          */
2310         if ((written == -EIOCBQUEUED) || (!ocfs2_iocb_is_rw_locked(iocb))) {
2311                 rw_level = -1;
2312         }
2313
2314         if (unlikely(written <= 0))
2315                 goto out;
2316
2317         if (((file->f_flags & O_DSYNC) && !direct_io) ||
2318             IS_SYNC(inode)) {
2319                 ret = filemap_fdatawrite_range(file->f_mapping,
2320                                                iocb->ki_pos - written,
2321                                                iocb->ki_pos - 1);
2322                 if (ret < 0)
2323                         written = ret;
2324
2325                 if (!ret) {
2326                         ret = jbd2_journal_force_commit(osb->journal->j_journal);
2327                         if (ret < 0)
2328                                 written = ret;
2329                 }
2330
2331                 if (!ret)
2332                         ret = filemap_fdatawait_range(file->f_mapping,
2333                                                       iocb->ki_pos - written,
2334                                                       iocb->ki_pos - 1);
2335         }
2336
2337 out:
2338         if (saved_ki_complete)
2339                 xchg(&iocb->ki_complete, saved_ki_complete);
2340
2341         if (rw_level != -1)
2342                 ocfs2_rw_unlock(inode, rw_level);
2343
2344 out_mutex:
2345         inode_unlock(inode);
2346
2347         if (written)
2348                 ret = written;
2349         return ret;
2350 }
2351
2352 static ssize_t ocfs2_file_read_iter(struct kiocb *iocb,
2353                                    struct iov_iter *to)
2354 {
2355         int ret = 0, rw_level = -1, lock_level = 0;
2356         struct file *filp = iocb->ki_filp;
2357         struct inode *inode = file_inode(filp);
2358
2359         trace_ocfs2_file_aio_read(inode, filp, filp->f_path.dentry,
2360                         (unsigned long long)OCFS2_I(inode)->ip_blkno,
2361                         filp->f_path.dentry->d_name.len,
2362                         filp->f_path.dentry->d_name.name,
2363                         to->nr_segs);   /* GRRRRR */
2364
2365
2366         if (!inode) {
2367                 ret = -EINVAL;
2368                 mlog_errno(ret);
2369                 goto bail;
2370         }
2371
2372         /*
2373          * buffered reads protect themselves in ->readpage().  O_DIRECT reads
2374          * need locks to protect pending reads from racing with truncate.
2375          */
2376         if (iocb->ki_flags & IOCB_DIRECT) {
2377                 ret = ocfs2_rw_lock(inode, 0);
2378                 if (ret < 0) {
2379                         mlog_errno(ret);
2380                         goto bail;
2381                 }
2382                 rw_level = 0;
2383                 /* communicate with ocfs2_dio_end_io */
2384                 ocfs2_iocb_set_rw_locked(iocb, rw_level);
2385         }
2386
2387         /*
2388          * We're fine letting folks race truncates and extending
2389          * writes with read across the cluster, just like they can
2390          * locally. Hence no rw_lock during read.
2391          *
2392          * Take and drop the meta data lock to update inode fields
2393          * like i_size. This allows the checks down below
2394          * generic_file_aio_read() a chance of actually working.
2395          */
2396         ret = ocfs2_inode_lock_atime(inode, filp->f_path.mnt, &lock_level);
2397         if (ret < 0) {
2398                 mlog_errno(ret);
2399                 goto bail;
2400         }
2401         ocfs2_inode_unlock(inode, lock_level);
2402
2403         ret = generic_file_read_iter(iocb, to);
2404         trace_generic_file_aio_read_ret(ret);
2405
2406         /* buffered aio wouldn't have proper lock coverage today */
2407         BUG_ON(ret == -EIOCBQUEUED && !(iocb->ki_flags & IOCB_DIRECT));
2408
2409         /* see ocfs2_file_write_iter */
2410         if (ret == -EIOCBQUEUED || !ocfs2_iocb_is_rw_locked(iocb)) {
2411                 rw_level = -1;
2412         }
2413
2414 bail:
2415         if (rw_level != -1)
2416                 ocfs2_rw_unlock(inode, rw_level);
2417
2418         return ret;
2419 }
2420
2421 /* Refer generic_file_llseek_unlocked() */
2422 static loff_t ocfs2_file_llseek(struct file *file, loff_t offset, int whence)
2423 {
2424         struct inode *inode = file->f_mapping->host;
2425         int ret = 0;
2426
2427         inode_lock(inode);
2428
2429         switch (whence) {
2430         case SEEK_SET:
2431                 break;
2432         case SEEK_END:
2433                 /* SEEK_END requires the OCFS2 inode lock for the file
2434                  * because it references the file's size.
2435                  */
2436                 ret = ocfs2_inode_lock(inode, NULL, 0);
2437                 if (ret < 0) {
2438                         mlog_errno(ret);
2439                         goto out;
2440                 }
2441                 offset += i_size_read(inode);
2442                 ocfs2_inode_unlock(inode, 0);
2443                 break;
2444         case SEEK_CUR:
2445                 if (offset == 0) {
2446                         offset = file->f_pos;
2447                         goto out;
2448                 }
2449                 offset += file->f_pos;
2450                 break;
2451         case SEEK_DATA:
2452         case SEEK_HOLE:
2453                 ret = ocfs2_seek_data_hole_offset(file, &offset, whence);
2454                 if (ret)
2455                         goto out;
2456                 break;
2457         default:
2458                 ret = -EINVAL;
2459                 goto out;
2460         }
2461
2462         offset = vfs_setpos(file, offset, inode->i_sb->s_maxbytes);
2463
2464 out:
2465         inode_unlock(inode);
2466         if (ret)
2467                 return ret;
2468         return offset;
2469 }
2470
2471 static int ocfs2_file_clone_range(struct file *file_in,
2472                                   loff_t pos_in,
2473                                   struct file *file_out,
2474                                   loff_t pos_out,
2475                                   u64 len)
2476 {
2477         return ocfs2_reflink_remap_range(file_in, pos_in, file_out, pos_out,
2478                                          len, false);
2479 }
2480
2481 static ssize_t ocfs2_file_dedupe_range(struct file *src_file,
2482                                        u64 loff,
2483                                        u64 len,
2484                                        struct file *dst_file,
2485                                        u64 dst_loff)
2486 {
2487         int error;
2488
2489         error = ocfs2_reflink_remap_range(src_file, loff, dst_file, dst_loff,
2490                                           len, true);
2491         if (error)
2492                 return error;
2493         return len;
2494 }
2495
2496 const struct inode_operations ocfs2_file_iops = {
2497         .setattr        = ocfs2_setattr,
2498         .getattr        = ocfs2_getattr,
2499         .permission     = ocfs2_permission,
2500         .listxattr      = ocfs2_listxattr,
2501         .fiemap         = ocfs2_fiemap,
2502         .get_acl        = ocfs2_iop_get_acl,
2503         .set_acl        = ocfs2_iop_set_acl,
2504 };
2505
2506 const struct inode_operations ocfs2_special_file_iops = {
2507         .setattr        = ocfs2_setattr,
2508         .getattr        = ocfs2_getattr,
2509         .permission     = ocfs2_permission,
2510         .get_acl        = ocfs2_iop_get_acl,
2511         .set_acl        = ocfs2_iop_set_acl,
2512 };
2513
2514 /*
2515  * Other than ->lock, keep ocfs2_fops and ocfs2_dops in sync with
2516  * ocfs2_fops_no_plocks and ocfs2_dops_no_plocks!
2517  */
2518 const struct file_operations ocfs2_fops = {
2519         .llseek         = ocfs2_file_llseek,
2520         .mmap           = ocfs2_mmap,
2521         .fsync          = ocfs2_sync_file,
2522         .release        = ocfs2_file_release,
2523         .open           = ocfs2_file_open,
2524         .read_iter      = ocfs2_file_read_iter,
2525         .write_iter     = ocfs2_file_write_iter,
2526         .unlocked_ioctl = ocfs2_ioctl,
2527 #ifdef CONFIG_COMPAT
2528         .compat_ioctl   = ocfs2_compat_ioctl,
2529 #endif
2530         .lock           = ocfs2_lock,
2531         .flock          = ocfs2_flock,
2532         .splice_read    = generic_file_splice_read,
2533         .splice_write   = iter_file_splice_write,
2534         .fallocate      = ocfs2_fallocate,
2535         .clone_file_range = ocfs2_file_clone_range,
2536         .dedupe_file_range = ocfs2_file_dedupe_range,
2537 };
2538
2539 const struct file_operations ocfs2_dops = {
2540         .llseek         = generic_file_llseek,
2541         .read           = generic_read_dir,
2542         .iterate        = ocfs2_readdir,
2543         .fsync          = ocfs2_sync_file,
2544         .release        = ocfs2_dir_release,
2545         .open           = ocfs2_dir_open,
2546         .unlocked_ioctl = ocfs2_ioctl,
2547 #ifdef CONFIG_COMPAT
2548         .compat_ioctl   = ocfs2_compat_ioctl,
2549 #endif
2550         .lock           = ocfs2_lock,
2551         .flock          = ocfs2_flock,
2552 };
2553
2554 /*
2555  * POSIX-lockless variants of our file_operations.
2556  *
2557  * These will be used if the underlying cluster stack does not support
2558  * posix file locking, if the user passes the "localflocks" mount
2559  * option, or if we have a local-only fs.
2560  *
2561  * ocfs2_flock is in here because all stacks handle UNIX file locks,
2562  * so we still want it in the case of no stack support for
2563  * plocks. Internally, it will do the right thing when asked to ignore
2564  * the cluster.
2565  */
2566 const struct file_operations ocfs2_fops_no_plocks = {
2567         .llseek         = ocfs2_file_llseek,
2568         .mmap           = ocfs2_mmap,
2569         .fsync          = ocfs2_sync_file,
2570         .release        = ocfs2_file_release,
2571         .open           = ocfs2_file_open,
2572         .read_iter      = ocfs2_file_read_iter,
2573         .write_iter     = ocfs2_file_write_iter,
2574         .unlocked_ioctl = ocfs2_ioctl,
2575 #ifdef CONFIG_COMPAT
2576         .compat_ioctl   = ocfs2_compat_ioctl,
2577 #endif
2578         .flock          = ocfs2_flock,
2579         .splice_read    = generic_file_splice_read,
2580         .splice_write   = iter_file_splice_write,
2581         .fallocate      = ocfs2_fallocate,
2582         .clone_file_range = ocfs2_file_clone_range,
2583         .dedupe_file_range = ocfs2_file_dedupe_range,
2584 };
2585
2586 const struct file_operations ocfs2_dops_no_plocks = {
2587         .llseek         = generic_file_llseek,
2588         .read           = generic_read_dir,
2589         .iterate        = ocfs2_readdir,
2590         .fsync          = ocfs2_sync_file,
2591         .release        = ocfs2_dir_release,
2592         .open           = ocfs2_dir_open,
2593         .unlocked_ioctl = ocfs2_ioctl,
2594 #ifdef CONFIG_COMPAT
2595         .compat_ioctl   = ocfs2_compat_ioctl,
2596 #endif
2597         .flock          = ocfs2_flock,
2598 };