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
  40 #define MLOG_MASK_PREFIX ML_INODE
  41 #include <cluster/masklog.h>
  42
  43 #include "ocfs2.h"
  44
  45 #include "alloc.h"
  46 #include "aops.h"
  47 #include "dir.h"
  48 #include "dlmglue.h"
  49 #include "extent_map.h"
  50 #include "file.h"
  51 #include "sysfile.h"
  52 #include "inode.h"
  53 #include "ioctl.h"
  54 #include "journal.h"
  55 #include "locks.h"
  56 #include "mmap.h"
  57 #include "suballoc.h"
  58 #include "super.h"
  59 #include "xattr.h"
  60 #include "acl.h"
  61 #include "quota.h"
  62 #include "refcounttree.h"
  63
  64 #include "buffer_head_io.h"
  65
  66 static int ocfs2_sync_inode(struct inode *inode)
  67 {
  68         filemap_fdatawrite(inode->i_mapping);
  69         return sync_mapping_buffers(inode->i_mapping);
  70 }
  71
  72 static int ocfs2_init_file_private(struct inode *inode, struct file *file)
  73 {
  74         struct ocfs2_file_private *fp;
  75
  76         fp = kzalloc(sizeof(struct ocfs2_file_private), GFP_KERNEL);
  77         if (!fp)
  78                 return -ENOMEM;
  79
  80         fp->fp_file = file;
  81         mutex_init(&fp->fp_mutex);
  82         ocfs2_file_lock_res_init(&fp->fp_flock, fp);
  83         file->private_data = fp;
  84
  85         return 0;
  86 }
  87
  88 static void ocfs2_free_file_private(struct inode *inode, struct file *file)
  89 {
  90         struct ocfs2_file_private *fp = file->private_data;
  91         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
  92
  93         if (fp) {
  94                 ocfs2_simple_drop_lockres(osb, &fp->fp_flock);
  95                 ocfs2_lock_res_free(&fp->fp_flock);
  96                 kfree(fp);
  97                 file->private_data = NULL;
  98         }
  99 }
 100
 101 static int ocfs2_file_open(struct inode *inode, struct file *file)
 102 {
 103         int status;
 104         int mode = file->f_flags;
 105         struct ocfs2_inode_info *oi = OCFS2_I(inode);
 106
 107         mlog_entry("(0x%p, 0x%p, '%.*s')\n", inode, file,
 108                    file->f_path.dentry->d_name.len, file->f_path.dentry->d_name.name);
 109
 110         if (file->f_mode & FMODE_WRITE)
 111                 dquot_initialize(inode);
 112
 113         spin_lock(&oi->ip_lock);
 114
 115         /* Check that the inode hasn't been wiped from disk by another
 116          * node. If it hasn't then we're safe as long as we hold the
 117          * spin lock until our increment of open count. */
 118         if (OCFS2_I(inode)->ip_flags & OCFS2_INODE_DELETED) {
 119                 spin_unlock(&oi->ip_lock);
 120
 121                 status = -ENOENT;
 122                 goto leave;
 123         }
 124
 125         if (mode & O_DIRECT)
 126                 oi->ip_flags |= OCFS2_INODE_OPEN_DIRECT;
 127
 128         oi->ip_open_count++;
 129         spin_unlock(&oi->ip_lock);
 130
 131         status = ocfs2_init_file_private(inode, file);
 132         if (status) {
 133                 /*
 134                  * We want to set open count back if we're failing the
 135                  * open.
 136                  */
 137                 spin_lock(&oi->ip_lock);
 138                 oi->ip_open_count--;
 139                 spin_unlock(&oi->ip_lock);
 140         }
 141
 142 leave:
 143         mlog_exit(status);
 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         mlog_entry("(0x%p, 0x%p, '%.*s')\n", inode, file,
 152                        file->f_path.dentry->d_name.len,
 153                        file->f_path.dentry->d_name.name);
 154
 155         spin_lock(&oi->ip_lock);
 156         if (!--oi->ip_open_count)
 157                 oi->ip_flags &= ~OCFS2_INODE_OPEN_DIRECT;
 158         spin_unlock(&oi->ip_lock);
 159
 160         ocfs2_free_file_private(inode, file);
 161
 162         mlog_exit(0);
 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, int datasync)
 179 {
 180         int err = 0;
 181         journal_t *journal;
 182         struct dentry *dentry = file->f_path.dentry;
 183         struct inode *inode = file->f_mapping->host;
 184         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 185
 186         mlog_entry("(0x%p, 0x%p, %d, '%.*s')\n", file, dentry, datasync,
 187                    dentry->d_name.len, dentry->d_name.name);
 188
 189         err = ocfs2_sync_inode(dentry->d_inode);
 190         if (err)
 191                 goto bail;
 192
 193         if (datasync && !(inode->i_state & I_DIRTY_DATASYNC))
 194                 goto bail;
 195
 196         journal = osb->journal->j_journal;
 197         err = jbd2_journal_force_commit(journal);
 198
 199 bail:
 200         mlog_exit(err);
 201
 202         return (err < 0) ? -EIO : 0;
 203 }
 204
 205 int ocfs2_should_update_atime(struct inode *inode,
 206                               struct vfsmount *vfsmnt)
 207 {
 208         struct timespec now;
 209         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 210
 211         if (ocfs2_is_hard_readonly(osb) || ocfs2_is_soft_readonly(osb))
 212                 return 0;
 213
 214         if ((inode->i_flags & S_NOATIME) ||
 215             ((inode->i_sb->s_flags & MS_NODIRATIME) && S_ISDIR(inode->i_mode)))
 216                 return 0;
 217
 218         /*
 219          * We can be called with no vfsmnt structure - NFSD will
 220          * sometimes do this.
 221          *
 222          * Note that our action here is different than touch_atime() -
 223          * if we can't tell whether this is a noatime mount, then we
 224          * don't know whether to trust the value of s_atime_quantum.
 225          */
 226         if (vfsmnt == NULL)
 227                 return 0;
 228
 229         if ((vfsmnt->mnt_flags & MNT_NOATIME) ||
 230             ((vfsmnt->mnt_flags & MNT_NODIRATIME) && S_ISDIR(inode->i_mode)))
 231                 return 0;
 232
 233         if (vfsmnt->mnt_flags & MNT_RELATIME) {
 234                 if ((timespec_compare(&inode->i_atime, &inode->i_mtime) <= 0) ||
 235                     (timespec_compare(&inode->i_atime, &inode->i_ctime) <= 0))
 236                         return 1;
 237
 238                 return 0;
 239         }
 240
 241         now = CURRENT_TIME;
 242         if ((now.tv_sec - inode->i_atime.tv_sec <= osb->s_atime_quantum))
 243                 return 0;
 244         else
 245                 return 1;
 246 }
 247
 248 int ocfs2_update_inode_atime(struct inode *inode,
 249                              struct buffer_head *bh)
 250 {
 251         int ret;
 252         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 253         handle_t *handle;
 254         struct ocfs2_dinode *di = (struct ocfs2_dinode *) bh->b_data;
 255
 256         mlog_entry_void();
 257
 258         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
 259         if (IS_ERR(handle)) {
 260                 ret = PTR_ERR(handle);
 261                 mlog_errno(ret);
 262                 goto out;
 263         }
 264
 265         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), bh,
 266                                       OCFS2_JOURNAL_ACCESS_WRITE);
 267         if (ret) {
 268                 mlog_errno(ret);
 269                 goto out_commit;
 270         }
 271
 272         /*
 273          * Don't use ocfs2_mark_inode_dirty() here as we don't always
 274          * have i_mutex to guard against concurrent changes to other
 275          * inode fields.
 276          */
 277         inode->i_atime = CURRENT_TIME;
 278         di->i_atime = cpu_to_le64(inode->i_atime.tv_sec);
 279         di->i_atime_nsec = cpu_to_le32(inode->i_atime.tv_nsec);
 280         ocfs2_journal_dirty(handle, bh);
 281
 282 out_commit:
 283         ocfs2_commit_trans(OCFS2_SB(inode->i_sb), handle);
 284 out:
 285         mlog_exit(ret);
 286         return ret;
 287 }
 288
 289 static int ocfs2_set_inode_size(handle_t *handle,
 290                                 struct inode *inode,
 291                                 struct buffer_head *fe_bh,
 292                                 u64 new_i_size)
 293 {
 294         int status;
 295
 296         mlog_entry_void();
 297         i_size_write(inode, new_i_size);
 298         inode->i_blocks = ocfs2_inode_sector_count(inode);
 299         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
 300
 301         status = ocfs2_mark_inode_dirty(handle, inode, fe_bh);
 302         if (status < 0) {
 303                 mlog_errno(status);
 304                 goto bail;
 305         }
 306
 307 bail:
 308         mlog_exit(status);
 309         return status;
 310 }
 311
 312 int ocfs2_simple_size_update(struct inode *inode,
 313                              struct buffer_head *di_bh,
 314                              u64 new_i_size)
 315 {
 316         int ret;
 317         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 318         handle_t *handle = NULL;
 319
 320         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
 321         if (IS_ERR(handle)) {
 322                 ret = PTR_ERR(handle);
 323                 mlog_errno(ret);
 324                 goto out;
 325         }
 326
 327         ret = ocfs2_set_inode_size(handle, inode, di_bh,
 328                                    new_i_size);
 329         if (ret < 0)
 330                 mlog_errno(ret);
 331
 332         ocfs2_commit_trans(osb, handle);
 333 out:
 334         return ret;
 335 }
 336
 337 static int ocfs2_cow_file_pos(struct inode *inode,
 338                               struct buffer_head *fe_bh,
 339                               u64 offset)
 340 {
 341         int status;
 342         u32 phys, cpos = offset >> OCFS2_SB(inode->i_sb)->s_clustersize_bits;
 343         unsigned int num_clusters = 0;
 344         unsigned int ext_flags = 0;
 345
 346         /*
 347          * If the new offset is aligned to the range of the cluster, there is
 348          * no space for ocfs2_zero_range_for_truncate to fill, so no need to
 349          * CoW either.
 350          */
 351         if ((offset & (OCFS2_SB(inode->i_sb)->s_clustersize - 1)) == 0)
 352                 return 0;
 353
 354         status = ocfs2_get_clusters(inode, cpos, &phys,
 355                                     &num_clusters, &ext_flags);
 356         if (status) {
 357                 mlog_errno(status);
 358                 goto out;
 359         }
 360
 361         if (!(ext_flags & OCFS2_EXT_REFCOUNTED))
 362                 goto out;
 363
 364         return ocfs2_refcount_cow(inode, fe_bh, cpos, 1, cpos+1);
 365
 366 out:
 367         return status;
 368 }
 369
 370 static int ocfs2_orphan_for_truncate(struct ocfs2_super *osb,
 371                                      struct inode *inode,
 372                                      struct buffer_head *fe_bh,
 373                                      u64 new_i_size)
 374 {
 375         int status;
 376         handle_t *handle;
 377         struct ocfs2_dinode *di;
 378         u64 cluster_bytes;
 379
 380         mlog_entry_void();
 381
 382         /*
 383          * We need to CoW the cluster contains the offset if it is reflinked
 384          * since we will call ocfs2_zero_range_for_truncate later which will
 385          * write "0" from offset to the end of the cluster.
 386          */
 387         status = ocfs2_cow_file_pos(inode, fe_bh, new_i_size);
 388         if (status) {
 389                 mlog_errno(status);
 390                 return status;
 391         }
 392
 393         /* TODO: This needs to actually orphan the inode in this
 394          * transaction. */
 395
 396         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
 397         if (IS_ERR(handle)) {
 398                 status = PTR_ERR(handle);
 399                 mlog_errno(status);
 400                 goto out;
 401         }
 402
 403         status = ocfs2_journal_access_di(handle, INODE_CACHE(inode), fe_bh,
 404                                          OCFS2_JOURNAL_ACCESS_WRITE);
 405         if (status < 0) {
 406                 mlog_errno(status);
 407                 goto out_commit;
 408         }
 409
 410         /*
 411          * Do this before setting i_size.
 412          */
 413         cluster_bytes = ocfs2_align_bytes_to_clusters(inode->i_sb, new_i_size);
 414         status = ocfs2_zero_range_for_truncate(inode, handle, new_i_size,
 415                                                cluster_bytes);
 416         if (status) {
 417                 mlog_errno(status);
 418                 goto out_commit;
 419         }
 420
 421         i_size_write(inode, new_i_size);
 422         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
 423
 424         di = (struct ocfs2_dinode *) fe_bh->b_data;
 425         di->i_size = cpu_to_le64(new_i_size);
 426         di->i_ctime = di->i_mtime = cpu_to_le64(inode->i_ctime.tv_sec);
 427         di->i_ctime_nsec = di->i_mtime_nsec = cpu_to_le32(inode->i_ctime.tv_nsec);
 428
 429         ocfs2_journal_dirty(handle, fe_bh);
 430
 431 out_commit:
 432         ocfs2_commit_trans(osb, handle);
 433 out:
 434
 435         mlog_exit(status);
 436         return status;
 437 }
 438
 439 static int ocfs2_truncate_file(struct inode *inode,
 440                                struct buffer_head *di_bh,
 441                                u64 new_i_size)
 442 {
 443         int status = 0;
 444         struct ocfs2_dinode *fe = NULL;
 445         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 446
 447         mlog_entry("(inode = %llu, new_i_size = %llu\n",
 448                    (unsigned long long)OCFS2_I(inode)->ip_blkno,
 449                    (unsigned long long)new_i_size);
 450
 451         /* We trust di_bh because it comes from ocfs2_inode_lock(), which
 452          * already validated it */
 453         fe = (struct ocfs2_dinode *) di_bh->b_data;
 454
 455         mlog_bug_on_msg(le64_to_cpu(fe->i_size) != i_size_read(inode),
 456                         "Inode %llu, inode i_size = %lld != di "
 457                         "i_size = %llu, i_flags = 0x%x\n",
 458                         (unsigned long long)OCFS2_I(inode)->ip_blkno,
 459                         i_size_read(inode),
 460                         (unsigned long long)le64_to_cpu(fe->i_size),
 461                         le32_to_cpu(fe->i_flags));
 462
 463         if (new_i_size > le64_to_cpu(fe->i_size)) {
 464                 mlog(0, "asked to truncate file with size (%llu) to size (%llu)!\n",
 465                      (unsigned long long)le64_to_cpu(fe->i_size),
 466                      (unsigned long long)new_i_size);
 467                 status = -EINVAL;
 468                 mlog_errno(status);
 469                 goto bail;
 470         }
 471
 472         mlog(0, "inode %llu, i_size = %llu, new_i_size = %llu\n",
 473              (unsigned long long)le64_to_cpu(fe->i_blkno),
 474              (unsigned long long)le64_to_cpu(fe->i_size),
 475              (unsigned long long)new_i_size);
 476
 477         /* lets handle the simple truncate cases before doing any more
 478          * cluster locking. */
 479         if (new_i_size == le64_to_cpu(fe->i_size))
 480                 goto bail;
 481
 482         down_write(&OCFS2_I(inode)->ip_alloc_sem);
 483
 484         ocfs2_resv_discard(&osb->osb_la_resmap,
 485                            &OCFS2_I(inode)->ip_la_data_resv);
 486
 487         /*
 488          * The inode lock forced other nodes to sync and drop their
 489          * pages, which (correctly) happens even if we have a truncate
 490          * without allocation change - ocfs2 cluster sizes can be much
 491          * greater than page size, so we have to truncate them
 492          * anyway.
 493          */
 494         unmap_mapping_range(inode->i_mapping, new_i_size + PAGE_SIZE - 1, 0, 1);
 495         truncate_inode_pages(inode->i_mapping, new_i_size);
 496
 497         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
 498                 status = ocfs2_truncate_inline(inode, di_bh, new_i_size,
 499                                                i_size_read(inode), 1);
 500                 if (status)
 501                         mlog_errno(status);
 502
 503                 goto bail_unlock_sem;
 504         }
 505
 506         /* alright, we're going to need to do a full blown alloc size
 507          * change. Orphan the inode so that recovery can complete the
 508          * truncate if necessary. This does the task of marking
 509          * i_size. */
 510         status = ocfs2_orphan_for_truncate(osb, inode, di_bh, new_i_size);
 511         if (status < 0) {
 512                 mlog_errno(status);
 513                 goto bail_unlock_sem;
 514         }
 515
 516         status = ocfs2_commit_truncate(osb, inode, di_bh);
 517         if (status < 0) {
 518                 mlog_errno(status);
 519                 goto bail_unlock_sem;
 520         }
 521
 522         /* TODO: orphan dir cleanup here. */
 523 bail_unlock_sem:
 524         up_write(&OCFS2_I(inode)->ip_alloc_sem);
 525
 526 bail:
 527         if (!status && OCFS2_I(inode)->ip_clusters == 0)
 528                 status = ocfs2_try_remove_refcount_tree(inode, di_bh);
 529
 530         mlog_exit(status);
 531         return status;
 532 }
 533
 534 /*
 535  * extend file allocation only here.
 536  * we'll update all the disk stuff, and oip->alloc_size
 537  *
 538  * expect stuff to be locked, a transaction started and enough data /
 539  * metadata reservations in the contexts.
 540  *
 541  * Will return -EAGAIN, and a reason if a restart is needed.
 542  * If passed in, *reason will always be set, even in error.
 543  */
 544 int ocfs2_add_inode_data(struct ocfs2_super *osb,
 545                          struct inode *inode,
 546                          u32 *logical_offset,
 547                          u32 clusters_to_add,
 548                          int mark_unwritten,
 549                          struct buffer_head *fe_bh,
 550                          handle_t *handle,
 551                          struct ocfs2_alloc_context *data_ac,
 552                          struct ocfs2_alloc_context *meta_ac,
 553                          enum ocfs2_alloc_restarted *reason_ret)
 554 {
 555         int ret;
 556         struct ocfs2_extent_tree et;
 557
 558         ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), fe_bh);
 559         ret = ocfs2_add_clusters_in_btree(handle, &et, logical_offset,
 560                                           clusters_to_add, mark_unwritten,
 561                                           data_ac, meta_ac, reason_ret);
 562
 563         return ret;
 564 }
 565
 566 static int __ocfs2_extend_allocation(struct inode *inode, u32 logical_start,
 567                                      u32 clusters_to_add, int mark_unwritten)
 568 {
 569         int status = 0;
 570         int restart_func = 0;
 571         int credits;
 572         u32 prev_clusters;
 573         struct buffer_head *bh = NULL;
 574         struct ocfs2_dinode *fe = NULL;
 575         handle_t *handle = NULL;
 576         struct ocfs2_alloc_context *data_ac = NULL;
 577         struct ocfs2_alloc_context *meta_ac = NULL;
 578         enum ocfs2_alloc_restarted why;
 579         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 580         struct ocfs2_extent_tree et;
 581         int did_quota = 0;
 582
 583         mlog_entry("(clusters_to_add = %u)\n", clusters_to_add);
 584
 585         /*
 586          * This function only exists for file systems which don't
 587          * support holes.
 588          */
 589         BUG_ON(mark_unwritten && !ocfs2_sparse_alloc(osb));
 590
 591         status = ocfs2_read_inode_block(inode, &bh);
 592         if (status < 0) {
 593                 mlog_errno(status);
 594                 goto leave;
 595         }
 596         fe = (struct ocfs2_dinode *) bh->b_data;
 597
 598 restart_all:
 599         BUG_ON(le32_to_cpu(fe->i_clusters) != OCFS2_I(inode)->ip_clusters);
 600
 601         mlog(0, "extend inode %llu, i_size = %lld, di->i_clusters = %u, "
 602              "clusters_to_add = %u\n",
 603              (unsigned long long)OCFS2_I(inode)->ip_blkno,
 604              (long long)i_size_read(inode), le32_to_cpu(fe->i_clusters),
 605              clusters_to_add);
 606         ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), bh);
 607         status = ocfs2_lock_allocators(inode, &et, clusters_to_add, 0,
 608                                        &data_ac, &meta_ac);
 609         if (status) {
 610                 mlog_errno(status);
 611                 goto leave;
 612         }
 613
 614         credits = ocfs2_calc_extend_credits(osb->sb, &fe->id2.i_list,
 615                                             clusters_to_add);
 616         handle = ocfs2_start_trans(osb, credits);
 617         if (IS_ERR(handle)) {
 618                 status = PTR_ERR(handle);
 619                 handle = NULL;
 620                 mlog_errno(status);
 621                 goto leave;
 622         }
 623
 624 restarted_transaction:
 625         status = dquot_alloc_space_nodirty(inode,
 626                         ocfs2_clusters_to_bytes(osb->sb, clusters_to_add));
 627         if (status)
 628                 goto leave;
 629         did_quota = 1;
 630
 631         /* reserve a write to the file entry early on - that we if we
 632          * run out of credits in the allocation path, we can still
 633          * update i_size. */
 634         status = ocfs2_journal_access_di(handle, INODE_CACHE(inode), bh,
 635                                          OCFS2_JOURNAL_ACCESS_WRITE);
 636         if (status < 0) {
 637                 mlog_errno(status);
 638                 goto leave;
 639         }
 640
 641         prev_clusters = OCFS2_I(inode)->ip_clusters;
 642
 643         status = ocfs2_add_inode_data(osb,
 644                                       inode,
 645                                       &logical_start,
 646                                       clusters_to_add,
 647                                       mark_unwritten,
 648                                       bh,
 649                                       handle,
 650                                       data_ac,
 651                                       meta_ac,
 652                                       &why);
 653         if ((status < 0) && (status != -EAGAIN)) {
 654                 if (status != -ENOSPC)
 655                         mlog_errno(status);
 656                 goto leave;
 657         }
 658
 659         ocfs2_journal_dirty(handle, bh);
 660
 661         spin_lock(&OCFS2_I(inode)->ip_lock);
 662         clusters_to_add -= (OCFS2_I(inode)->ip_clusters - prev_clusters);
 663         spin_unlock(&OCFS2_I(inode)->ip_lock);
 664         /* Release unused quota reservation */
 665         dquot_free_space(inode,
 666                         ocfs2_clusters_to_bytes(osb->sb, clusters_to_add));
 667         did_quota = 0;
 668
 669         if (why != RESTART_NONE && clusters_to_add) {
 670                 if (why == RESTART_META) {
 671                         mlog(0, "restarting function.\n");
 672                         restart_func = 1;
 673                         status = 0;
 674                 } else {
 675                         BUG_ON(why != RESTART_TRANS);
 676
 677                         mlog(0, "restarting transaction.\n");
 678                         /* TODO: This can be more intelligent. */
 679                         credits = ocfs2_calc_extend_credits(osb->sb,
 680                                                             &fe->id2.i_list,
 681                                                             clusters_to_add);
 682                         status = ocfs2_extend_trans(handle, credits);
 683                         if (status < 0) {
 684                                 /* handle still has to be committed at
 685                                  * this point. */
 686                                 status = -ENOMEM;
 687                                 mlog_errno(status);
 688                                 goto leave;
 689                         }
 690                         goto restarted_transaction;
 691                 }
 692         }
 693
 694         mlog(0, "fe: i_clusters = %u, i_size=%llu\n",
 695              le32_to_cpu(fe->i_clusters),
 696              (unsigned long long)le64_to_cpu(fe->i_size));
 697         mlog(0, "inode: ip_clusters=%u, i_size=%lld\n",
 698              OCFS2_I(inode)->ip_clusters, (long long)i_size_read(inode));
 699
 700 leave:
 701         if (status < 0 && did_quota)
 702                 dquot_free_space(inode,
 703                         ocfs2_clusters_to_bytes(osb->sb, clusters_to_add));
 704         if (handle) {
 705                 ocfs2_commit_trans(osb, handle);
 706                 handle = NULL;
 707         }
 708         if (data_ac) {
 709                 ocfs2_free_alloc_context(data_ac);
 710                 data_ac = NULL;
 711         }
 712         if (meta_ac) {
 713                 ocfs2_free_alloc_context(meta_ac);
 714                 meta_ac = NULL;
 715         }
 716         if ((!status) && restart_func) {
 717                 restart_func = 0;
 718                 goto restart_all;
 719         }
 720         brelse(bh);
 721         bh = NULL;
 722
 723         mlog_exit(status);
 724         return status;
 725 }
 726
 727 /* Some parts of this taken from generic_cont_expand, which turned out
 728  * to be too fragile to do exactly what we need without us having to
 729  * worry about recursive locking in ->write_begin() and ->write_end(). */
 730 static int ocfs2_write_zero_page(struct inode *inode,
 731                                  u64 size)
 732 {
 733         struct address_space *mapping = inode->i_mapping;
 734         struct page *page;
 735         unsigned long index;
 736         unsigned int offset;
 737         handle_t *handle = NULL;
 738         int ret;
 739
 740         offset = (size & (PAGE_CACHE_SIZE-1)); /* Within page */
 741         /* ugh.  in prepare/commit_write, if from==to==start of block, we
 742         ** skip the prepare.  make sure we never send an offset for the start
 743         ** of a block
 744         */
 745         if ((offset & (inode->i_sb->s_blocksize - 1)) == 0) {
 746                 offset++;
 747         }
 748         index = size >> PAGE_CACHE_SHIFT;
 749
 750         page = grab_cache_page(mapping, index);
 751         if (!page) {
 752                 ret = -ENOMEM;
 753                 mlog_errno(ret);
 754                 goto out;
 755         }
 756
 757         ret = ocfs2_prepare_write_nolock(inode, page, offset, offset);
 758         if (ret < 0) {
 759                 mlog_errno(ret);
 760                 goto out_unlock;
 761         }
 762
 763         if (ocfs2_should_order_data(inode)) {
 764                 handle = ocfs2_start_walk_page_trans(inode, page, offset,
 765                                                      offset);
 766                 if (IS_ERR(handle)) {
 767                         ret = PTR_ERR(handle);
 768                         handle = NULL;
 769                         goto out_unlock;
 770                 }
 771         }
 772
 773         /* must not update i_size! */
 774         ret = block_commit_write(page, offset, offset);
 775         if (ret < 0)
 776                 mlog_errno(ret);
 777         else
 778                 ret = 0;
 779
 780         if (handle)
 781                 ocfs2_commit_trans(OCFS2_SB(inode->i_sb), handle);
 782 out_unlock:
 783         unlock_page(page);
 784         page_cache_release(page);
 785 out:
 786         return ret;
 787 }
 788
 789 static int ocfs2_zero_extend(struct inode *inode,
 790                              u64 zero_to_size)
 791 {
 792         int ret = 0;
 793         u64 start_off;
 794         struct super_block *sb = inode->i_sb;
 795
 796         start_off = ocfs2_align_bytes_to_blocks(sb, i_size_read(inode));
 797         while (start_off < zero_to_size) {
 798                 ret = ocfs2_write_zero_page(inode, start_off);
 799                 if (ret < 0) {
 800                         mlog_errno(ret);
 801                         goto out;
 802                 }
 803
 804                 start_off += sb->s_blocksize;
 805
 806                 /*
 807                  * Very large extends have the potential to lock up
 808                  * the cpu for extended periods of time.
 809                  */
 810                 cond_resched();
 811         }
 812
 813 out:
 814         return ret;
 815 }
 816
 817 int ocfs2_extend_no_holes(struct inode *inode, u64 new_i_size, u64 zero_to)
 818 {
 819         int ret;
 820         u32 clusters_to_add;
 821         struct ocfs2_inode_info *oi = OCFS2_I(inode);
 822
 823         clusters_to_add = ocfs2_clusters_for_bytes(inode->i_sb, new_i_size);
 824         if (clusters_to_add < oi->ip_clusters)
 825                 clusters_to_add = 0;
 826         else
 827                 clusters_to_add -= oi->ip_clusters;
 828
 829         if (clusters_to_add) {
 830                 ret = __ocfs2_extend_allocation(inode, oi->ip_clusters,
 831                                                 clusters_to_add, 0);
 832                 if (ret) {
 833                         mlog_errno(ret);
 834                         goto out;
 835                 }
 836         }
 837
 838         /*
 839          * Call this even if we don't add any clusters to the tree. We
 840          * still need to zero the area between the old i_size and the
 841          * new i_size.
 842          */
 843         ret = ocfs2_zero_extend(inode, zero_to);
 844         if (ret < 0)
 845                 mlog_errno(ret);
 846
 847 out:
 848         return ret;
 849 }
 850
 851 static int ocfs2_extend_file(struct inode *inode,
 852                              struct buffer_head *di_bh,
 853                              u64 new_i_size)
 854 {
 855         int ret = 0;
 856         struct ocfs2_inode_info *oi = OCFS2_I(inode);
 857
 858         BUG_ON(!di_bh);
 859
 860         /* setattr sometimes calls us like this. */
 861         if (new_i_size == 0)
 862                 goto out;
 863
 864         if (i_size_read(inode) == new_i_size)
 865                 goto out;
 866         BUG_ON(new_i_size < i_size_read(inode));
 867
 868         /*
 869          * Fall through for converting inline data, even if the fs
 870          * supports sparse files.
 871          *
 872          * The check for inline data here is legal - nobody can add
 873          * the feature since we have i_mutex. We must check it again
 874          * after acquiring ip_alloc_sem though, as paths like mmap
 875          * might have raced us to converting the inode to extents.
 876          */
 877         if (!(oi->ip_dyn_features & OCFS2_INLINE_DATA_FL)
 878             && ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)))
 879                 goto out_update_size;
 880
 881         /*
 882          * The alloc sem blocks people in read/write from reading our
 883          * allocation until we're done changing it. We depend on
 884          * i_mutex to block other extend/truncate calls while we're
 885          * here.
 886          */
 887         down_write(&oi->ip_alloc_sem);
 888
 889         if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
 890                 /*
 891                  * We can optimize small extends by keeping the inodes
 892                  * inline data.
 893                  */
 894                 if (ocfs2_size_fits_inline_data(di_bh, new_i_size)) {
 895                         up_write(&oi->ip_alloc_sem);
 896                         goto out_update_size;
 897                 }
 898
 899                 ret = ocfs2_convert_inline_data_to_extents(inode, di_bh);
 900                 if (ret) {
 901                         up_write(&oi->ip_alloc_sem);
 902
 903                         mlog_errno(ret);
 904                         goto out;
 905                 }
 906         }
 907
 908         if (!ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)))
 909                 ret = ocfs2_extend_no_holes(inode, new_i_size, new_i_size);
 910
 911         up_write(&oi->ip_alloc_sem);
 912
 913         if (ret < 0) {
 914                 mlog_errno(ret);
 915                 goto out;
 916         }
 917
 918 out_update_size:
 919         ret = ocfs2_simple_size_update(inode, di_bh, new_i_size);
 920         if (ret < 0)
 921                 mlog_errno(ret);
 922
 923 out:
 924         return ret;
 925 }
 926
 927 int ocfs2_setattr(struct dentry *dentry, struct iattr *attr)
 928 {
 929         int status = 0, size_change;
 930         struct inode *inode = dentry->d_inode;
 931         struct super_block *sb = inode->i_sb;
 932         struct ocfs2_super *osb = OCFS2_SB(sb);
 933         struct buffer_head *bh = NULL;
 934         handle_t *handle = NULL;
 935         struct dquot *transfer_to[MAXQUOTAS] = { };
 936         int qtype;
 937
 938         mlog_entry("(0x%p, '%.*s')\n", dentry,
 939                    dentry->d_name.len, dentry->d_name.name);
 940
 941         /* ensuring we don't even attempt to truncate a symlink */
 942         if (S_ISLNK(inode->i_mode))
 943                 attr->ia_valid &= ~ATTR_SIZE;
 944
 945         if (attr->ia_valid & ATTR_MODE)
 946                 mlog(0, "mode change: %d\n", attr->ia_mode);
 947         if (attr->ia_valid & ATTR_UID)
 948                 mlog(0, "uid change: %d\n", attr->ia_uid);
 949         if (attr->ia_valid & ATTR_GID)
 950                 mlog(0, "gid change: %d\n", attr->ia_gid);
 951         if (attr->ia_valid & ATTR_SIZE)
 952                 mlog(0, "size change...\n");
 953         if (attr->ia_valid & (ATTR_ATIME | ATTR_MTIME | ATTR_CTIME))
 954                 mlog(0, "time change...\n");
 955
 956 #define OCFS2_VALID_ATTRS (ATTR_ATIME | ATTR_MTIME | ATTR_CTIME | ATTR_SIZE \
 957                            | ATTR_GID | ATTR_UID | ATTR_MODE)
 958         if (!(attr->ia_valid & OCFS2_VALID_ATTRS)) {
 959                 mlog(0, "can't handle attrs: 0x%x\n", attr->ia_valid);
 960                 return 0;
 961         }
 962
 963         status = inode_change_ok(inode, attr);
 964         if (status)
 965                 return status;
 966
 967         if (is_quota_modification(inode, attr))
 968                 dquot_initialize(inode);
 969         size_change = S_ISREG(inode->i_mode) && attr->ia_valid & ATTR_SIZE;
 970         if (size_change) {
 971                 status = ocfs2_rw_lock(inode, 1);
 972                 if (status < 0) {
 973                         mlog_errno(status);
 974                         goto bail;
 975                 }
 976         }
 977
 978         status = ocfs2_inode_lock(inode, &bh, 1);
 979         if (status < 0) {
 980                 if (status != -ENOENT)
 981                         mlog_errno(status);
 982                 goto bail_unlock_rw;
 983         }
 984
 985         if (size_change && attr->ia_size != i_size_read(inode)) {
 986                 status = inode_newsize_ok(inode, attr->ia_size);
 987                 if (status)
 988                         goto bail_unlock;
 989
 990                 if (i_size_read(inode) > attr->ia_size) {
 991                         if (ocfs2_should_order_data(inode)) {
 992                                 status = ocfs2_begin_ordered_truncate(inode,
 993                                                                       attr->ia_size);
 994                                 if (status)
 995                                         goto bail_unlock;
 996                         }
 997                         status = ocfs2_truncate_file(inode, bh, attr->ia_size);
 998                 } else
 999                         status = ocfs2_extend_file(inode, bh, attr->ia_size);
1000                 if (status < 0) {
1001                         if (status != -ENOSPC)
1002                                 mlog_errno(status);
1003                         status = -ENOSPC;
1004                         goto bail_unlock;
1005                 }
1006         }
1007
1008         if ((attr->ia_valid & ATTR_UID && attr->ia_uid != inode->i_uid) ||
1009             (attr->ia_valid & ATTR_GID && attr->ia_gid != inode->i_gid)) {
1010                 /*
1011                  * Gather pointers to quota structures so that allocation /
1012                  * freeing of quota structures happens here and not inside
1013                  * dquot_transfer() where we have problems with lock ordering
1014                  */
1015                 if (attr->ia_valid & ATTR_UID && attr->ia_uid != inode->i_uid
1016                     && OCFS2_HAS_RO_COMPAT_FEATURE(sb,
1017                     OCFS2_FEATURE_RO_COMPAT_USRQUOTA)) {
1018                         transfer_to[USRQUOTA] = dqget(sb, attr->ia_uid,
1019                                                       USRQUOTA);
1020                         if (!transfer_to[USRQUOTA]) {
1021                                 status = -ESRCH;
1022                                 goto bail_unlock;
1023                         }
1024                 }
1025                 if (attr->ia_valid & ATTR_GID && attr->ia_gid != inode->i_gid
1026                     && OCFS2_HAS_RO_COMPAT_FEATURE(sb,
1027                     OCFS2_FEATURE_RO_COMPAT_GRPQUOTA)) {
1028                         transfer_to[GRPQUOTA] = dqget(sb, attr->ia_gid,
1029                                                       GRPQUOTA);
1030                         if (!transfer_to[GRPQUOTA]) {
1031                                 status = -ESRCH;
1032                                 goto bail_unlock;
1033                         }
1034                 }
1035                 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS +
1036                                            2 * ocfs2_quota_trans_credits(sb));
1037                 if (IS_ERR(handle)) {
1038                         status = PTR_ERR(handle);
1039                         mlog_errno(status);
1040                         goto bail_unlock;
1041                 }
1042                 status = __dquot_transfer(inode, transfer_to);
1043                 if (status < 0)
1044                         goto bail_commit;
1045         } else {
1046                 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1047                 if (IS_ERR(handle)) {
1048                         status = PTR_ERR(handle);
1049                         mlog_errno(status);
1050                         goto bail_unlock;
1051                 }
1052         }
1053
1054         /*
1055          * This will intentionally not wind up calling simple_setsize(),
1056          * since all the work for a size change has been done above.
1057          * Otherwise, we could get into problems with truncate as
1058          * ip_alloc_sem is used there to protect against i_size
1059          * changes.
1060          */
1061         status = inode_setattr(inode, attr);
1062         if (status < 0) {
1063                 mlog_errno(status);
1064                 goto bail_commit;
1065         }
1066
1067         status = ocfs2_mark_inode_dirty(handle, inode, bh);
1068         if (status < 0)
1069                 mlog_errno(status);
1070
1071 bail_commit:
1072         ocfs2_commit_trans(osb, handle);
1073 bail_unlock:
1074         ocfs2_inode_unlock(inode, 1);
1075 bail_unlock_rw:
1076         if (size_change)
1077                 ocfs2_rw_unlock(inode, 1);
1078 bail:
1079         brelse(bh);
1080
1081         /* Release quota pointers in case we acquired them */
1082         for (qtype = 0; qtype < MAXQUOTAS; qtype++)
1083                 dqput(transfer_to[qtype]);
1084
1085         if (!status && attr->ia_valid & ATTR_MODE) {
1086                 status = ocfs2_acl_chmod(inode);
1087                 if (status < 0)
1088                         mlog_errno(status);
1089         }
1090
1091         mlog_exit(status);
1092         return status;
1093 }
1094
1095 int ocfs2_getattr(struct vfsmount *mnt,
1096                   struct dentry *dentry,
1097                   struct kstat *stat)
1098 {
1099         struct inode *inode = dentry->d_inode;
1100         struct super_block *sb = dentry->d_inode->i_sb;
1101         struct ocfs2_super *osb = sb->s_fs_info;
1102         int err;
1103
1104         mlog_entry_void();
1105
1106         err = ocfs2_inode_revalidate(dentry);
1107         if (err) {
1108                 if (err != -ENOENT)
1109                         mlog_errno(err);
1110                 goto bail;
1111         }
1112
1113         generic_fillattr(inode, stat);
1114
1115         /* We set the blksize from the cluster size for performance */
1116         stat->blksize = osb->s_clustersize;
1117
1118 bail:
1119         mlog_exit(err);
1120
1121         return err;
1122 }
1123
1124 int ocfs2_permission(struct inode *inode, int mask)
1125 {
1126         int ret;
1127
1128         mlog_entry_void();
1129
1130         ret = ocfs2_inode_lock(inode, NULL, 0);
1131         if (ret) {
1132                 if (ret != -ENOENT)
1133                         mlog_errno(ret);
1134                 goto out;
1135         }
1136
1137         ret = generic_permission(inode, mask, ocfs2_check_acl);
1138
1139         ocfs2_inode_unlock(inode, 0);
1140 out:
1141         mlog_exit(ret);
1142         return ret;
1143 }
1144
1145 static int __ocfs2_write_remove_suid(struct inode *inode,
1146                                      struct buffer_head *bh)
1147 {
1148         int ret;
1149         handle_t *handle;
1150         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1151         struct ocfs2_dinode *di;
1152
1153         mlog_entry("(Inode %llu, mode 0%o)\n",
1154                    (unsigned long long)OCFS2_I(inode)->ip_blkno, inode->i_mode);
1155
1156         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1157         if (IS_ERR(handle)) {
1158                 ret = PTR_ERR(handle);
1159                 mlog_errno(ret);
1160                 goto out;
1161         }
1162
1163         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), bh,
1164                                       OCFS2_JOURNAL_ACCESS_WRITE);
1165         if (ret < 0) {
1166                 mlog_errno(ret);
1167                 goto out_trans;
1168         }
1169
1170         inode->i_mode &= ~S_ISUID;
1171         if ((inode->i_mode & S_ISGID) && (inode->i_mode & S_IXGRP))
1172                 inode->i_mode &= ~S_ISGID;
1173
1174         di = (struct ocfs2_dinode *) bh->b_data;
1175         di->i_mode = cpu_to_le16(inode->i_mode);
1176
1177         ocfs2_journal_dirty(handle, bh);
1178
1179 out_trans:
1180         ocfs2_commit_trans(osb, handle);
1181 out:
1182         mlog_exit(ret);
1183         return ret;
1184 }
1185
1186 /*
1187  * Will look for holes and unwritten extents in the range starting at
1188  * pos for count bytes (inclusive).
1189  */
1190 static int ocfs2_check_range_for_holes(struct inode *inode, loff_t pos,
1191                                        size_t count)
1192 {
1193         int ret = 0;
1194         unsigned int extent_flags;
1195         u32 cpos, clusters, extent_len, phys_cpos;
1196         struct super_block *sb = inode->i_sb;
1197
1198         cpos = pos >> OCFS2_SB(sb)->s_clustersize_bits;
1199         clusters = ocfs2_clusters_for_bytes(sb, pos + count) - cpos;
1200
1201         while (clusters) {
1202                 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos, &extent_len,
1203                                          &extent_flags);
1204                 if (ret < 0) {
1205                         mlog_errno(ret);
1206                         goto out;
1207                 }
1208
1209                 if (phys_cpos == 0 || (extent_flags & OCFS2_EXT_UNWRITTEN)) {
1210                         ret = 1;
1211                         break;
1212                 }
1213
1214                 if (extent_len > clusters)
1215                         extent_len = clusters;
1216
1217                 clusters -= extent_len;
1218                 cpos += extent_len;
1219         }
1220 out:
1221         return ret;
1222 }
1223
1224 static int ocfs2_write_remove_suid(struct inode *inode)
1225 {
1226         int ret;
1227         struct buffer_head *bh = NULL;
1228
1229         ret = ocfs2_read_inode_block(inode, &bh);
1230         if (ret < 0) {
1231                 mlog_errno(ret);
1232                 goto out;
1233         }
1234
1235         ret =  __ocfs2_write_remove_suid(inode, bh);
1236 out:
1237         brelse(bh);
1238         return ret;
1239 }
1240
1241 /*
1242  * Allocate enough extents to cover the region starting at byte offset
1243  * start for len bytes. Existing extents are skipped, any extents
1244  * added are marked as "unwritten".
1245  */
1246 static int ocfs2_allocate_unwritten_extents(struct inode *inode,
1247                                             u64 start, u64 len)
1248 {
1249         int ret;
1250         u32 cpos, phys_cpos, clusters, alloc_size;
1251         u64 end = start + len;
1252         struct buffer_head *di_bh = NULL;
1253
1254         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1255                 ret = ocfs2_read_inode_block(inode, &di_bh);
1256                 if (ret) {
1257                         mlog_errno(ret);
1258                         goto out;
1259                 }
1260
1261                 /*
1262                  * Nothing to do if the requested reservation range
1263                  * fits within the inode.
1264                  */
1265                 if (ocfs2_size_fits_inline_data(di_bh, end))
1266                         goto out;
1267
1268                 ret = ocfs2_convert_inline_data_to_extents(inode, di_bh);
1269                 if (ret) {
1270                         mlog_errno(ret);
1271                         goto out;
1272                 }
1273         }
1274
1275         /*
1276          * We consider both start and len to be inclusive.
1277          */
1278         cpos = start >> OCFS2_SB(inode->i_sb)->s_clustersize_bits;
1279         clusters = ocfs2_clusters_for_bytes(inode->i_sb, start + len);
1280         clusters -= cpos;
1281
1282         while (clusters) {
1283                 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos,
1284                                          &alloc_size, NULL);
1285                 if (ret) {
1286                         mlog_errno(ret);
1287                         goto out;
1288                 }
1289
1290                 /*
1291                  * Hole or existing extent len can be arbitrary, so
1292                  * cap it to our own allocation request.
1293                  */
1294                 if (alloc_size > clusters)
1295                         alloc_size = clusters;
1296
1297                 if (phys_cpos) {
1298                         /*
1299                          * We already have an allocation at this
1300                          * region so we can safely skip it.
1301                          */
1302                         goto next;
1303                 }
1304
1305                 ret = __ocfs2_extend_allocation(inode, cpos, alloc_size, 1);
1306                 if (ret) {
1307                         if (ret != -ENOSPC)
1308                                 mlog_errno(ret);
1309                         goto out;
1310                 }
1311
1312 next:
1313                 cpos += alloc_size;
1314                 clusters -= alloc_size;
1315         }
1316
1317         ret = 0;
1318 out:
1319
1320         brelse(di_bh);
1321         return ret;
1322 }
1323
1324 /*
1325  * Truncate a byte range, avoiding pages within partial clusters. This
1326  * preserves those pages for the zeroing code to write to.
1327  */
1328 static void ocfs2_truncate_cluster_pages(struct inode *inode, u64 byte_start,
1329                                          u64 byte_len)
1330 {
1331         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1332         loff_t start, end;
1333         struct address_space *mapping = inode->i_mapping;
1334
1335         start = (loff_t)ocfs2_align_bytes_to_clusters(inode->i_sb, byte_start);
1336         end = byte_start + byte_len;
1337         end = end & ~(osb->s_clustersize - 1);
1338
1339         if (start < end) {
1340                 unmap_mapping_range(mapping, start, end - start, 0);
1341                 truncate_inode_pages_range(mapping, start, end - 1);
1342         }
1343 }
1344
1345 static int ocfs2_zero_partial_clusters(struct inode *inode,
1346                                        u64 start, u64 len)
1347 {
1348         int ret = 0;
1349         u64 tmpend, end = start + len;
1350         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1351         unsigned int csize = osb->s_clustersize;
1352         handle_t *handle;
1353
1354         /*
1355          * The "start" and "end" values are NOT necessarily part of
1356          * the range whose allocation is being deleted. Rather, this
1357          * is what the user passed in with the request. We must zero
1358          * partial clusters here. There's no need to worry about
1359          * physical allocation - the zeroing code knows to skip holes.
1360          */
1361         mlog(0, "byte start: %llu, end: %llu\n",
1362              (unsigned long long)start, (unsigned long long)end);
1363
1364         /*
1365          * If both edges are on a cluster boundary then there's no
1366          * zeroing required as the region is part of the allocation to
1367          * be truncated.
1368          */
1369         if ((start & (csize - 1)) == 0 && (end & (csize - 1)) == 0)
1370                 goto out;
1371
1372         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1373         if (IS_ERR(handle)) {
1374                 ret = PTR_ERR(handle);
1375                 mlog_errno(ret);
1376                 goto out;
1377         }
1378
1379         /*
1380          * We want to get the byte offset of the end of the 1st cluster.
1381          */
1382         tmpend = (u64)osb->s_clustersize + (start & ~(osb->s_clustersize - 1));
1383         if (tmpend > end)
1384                 tmpend = end;
1385
1386         mlog(0, "1st range: start: %llu, tmpend: %llu\n",
1387              (unsigned long long)start, (unsigned long long)tmpend);
1388
1389         ret = ocfs2_zero_range_for_truncate(inode, handle, start, tmpend);
1390         if (ret)
1391                 mlog_errno(ret);
1392
1393         if (tmpend < end) {
1394                 /*
1395                  * This may make start and end equal, but the zeroing
1396                  * code will skip any work in that case so there's no
1397                  * need to catch it up here.
1398                  */
1399                 start = end & ~(osb->s_clustersize - 1);
1400
1401                 mlog(0, "2nd range: start: %llu, end: %llu\n",
1402                      (unsigned long long)start, (unsigned long long)end);
1403
1404                 ret = ocfs2_zero_range_for_truncate(inode, handle, start, end);
1405                 if (ret)
1406                         mlog_errno(ret);
1407         }
1408
1409         ocfs2_commit_trans(osb, handle);
1410 out:
1411         return ret;
1412 }
1413
1414 static int ocfs2_find_rec(struct ocfs2_extent_list *el, u32 pos)
1415 {
1416         int i;
1417         struct ocfs2_extent_rec *rec = NULL;
1418
1419         for (i = le16_to_cpu(el->l_next_free_rec) - 1; i >= 0; i--) {
1420
1421                 rec = &el->l_recs[i];
1422
1423                 if (le32_to_cpu(rec->e_cpos) < pos)
1424                         break;
1425         }
1426
1427         return i;
1428 }
1429
1430 /*
1431  * Helper to calculate the punching pos and length in one run, we handle the
1432  * following three cases in order:
1433  *
1434  * - remove the entire record
1435  * - remove a partial record
1436  * - no record needs to be removed (hole-punching completed)
1437 */
1438 static void ocfs2_calc_trunc_pos(struct inode *inode,
1439                                  struct ocfs2_extent_list *el,
1440                                  struct ocfs2_extent_rec *rec,
1441                                  u32 trunc_start, u32 *trunc_cpos,
1442                                  u32 *trunc_len, u32 *trunc_end,
1443                                  u64 *blkno, int *done)
1444 {
1445         int ret = 0;
1446         u32 coff, range;
1447
1448         range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
1449
1450         if (le32_to_cpu(rec->e_cpos) >= trunc_start) {
1451                 *trunc_cpos = le32_to_cpu(rec->e_cpos);
1452                 /*
1453                  * Skip holes if any.
1454                  */
1455                 if (range < *trunc_end)
1456                         *trunc_end = range;
1457                 *trunc_len = *trunc_end - le32_to_cpu(rec->e_cpos);
1458                 *blkno = le64_to_cpu(rec->e_blkno);
1459                 *trunc_end = le32_to_cpu(rec->e_cpos);
1460         } else if (range > trunc_start) {
1461                 *trunc_cpos = trunc_start;
1462                 *trunc_len = *trunc_end - trunc_start;
1463                 coff = trunc_start - le32_to_cpu(rec->e_cpos);
1464                 *blkno = le64_to_cpu(rec->e_blkno) +
1465                                 ocfs2_clusters_to_blocks(inode->i_sb, coff);
1466                 *trunc_end = trunc_start;
1467         } else {
1468                 /*
1469                  * It may have two following possibilities:
1470                  *
1471                  * - last record has been removed
1472                  * - trunc_start was within a hole
1473                  *
1474                  * both two cases mean the completion of hole punching.
1475                  */
1476                 ret = 1;
1477         }
1478
1479         *done = ret;
1480 }
1481
1482 static int ocfs2_remove_inode_range(struct inode *inode,
1483                                     struct buffer_head *di_bh, u64 byte_start,
1484                                     u64 byte_len)
1485 {
1486         int ret = 0, flags = 0, done = 0, i;
1487         u32 trunc_start, trunc_len, trunc_end, trunc_cpos, phys_cpos;
1488         u32 cluster_in_el;
1489         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1490         struct ocfs2_cached_dealloc_ctxt dealloc;
1491         struct address_space *mapping = inode->i_mapping;
1492         struct ocfs2_extent_tree et;
1493         struct ocfs2_path *path = NULL;
1494         struct ocfs2_extent_list *el = NULL;
1495         struct ocfs2_extent_rec *rec = NULL;
1496         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
1497         u64 blkno, refcount_loc = le64_to_cpu(di->i_refcount_loc);
1498
1499         ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), di_bh);
1500         ocfs2_init_dealloc_ctxt(&dealloc);
1501
1502         if (byte_len == 0)
1503                 return 0;
1504
1505         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1506                 ret = ocfs2_truncate_inline(inode, di_bh, byte_start,
1507                                             byte_start + byte_len, 0);
1508                 if (ret) {
1509                         mlog_errno(ret);
1510                         goto out;
1511                 }
1512                 /*
1513                  * There's no need to get fancy with the page cache
1514                  * truncate of an inline-data inode. We're talking
1515                  * about less than a page here, which will be cached
1516                  * in the dinode buffer anyway.
1517                  */
1518                 unmap_mapping_range(mapping, 0, 0, 0);
1519                 truncate_inode_pages(mapping, 0);
1520                 goto out;
1521         }
1522
1523         /*
1524          * For reflinks, we may need to CoW 2 clusters which might be
1525          * partially zero'd later, if hole's start and end offset were
1526          * within one cluster(means is not exactly aligned to clustersize).
1527          */
1528
1529         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL) {
1530
1531                 ret = ocfs2_cow_file_pos(inode, di_bh, byte_start);
1532                 if (ret) {
1533                         mlog_errno(ret);
1534                         goto out;
1535                 }
1536
1537                 ret = ocfs2_cow_file_pos(inode, di_bh, byte_start + byte_len);
1538                 if (ret) {
1539                         mlog_errno(ret);
1540                         goto out;
1541                 }
1542         }
1543
1544         trunc_start = ocfs2_clusters_for_bytes(osb->sb, byte_start);
1545         trunc_end = (byte_start + byte_len) >> osb->s_clustersize_bits;
1546         cluster_in_el = trunc_end;
1547
1548         mlog(0, "Inode: %llu, start: %llu, len: %llu, cstart: %u, cend: %u\n",
1549              (unsigned long long)OCFS2_I(inode)->ip_blkno,
1550              (unsigned long long)byte_start,
1551              (unsigned long long)byte_len, trunc_start, trunc_end);
1552
1553         ret = ocfs2_zero_partial_clusters(inode, byte_start, byte_len);
1554         if (ret) {
1555                 mlog_errno(ret);
1556                 goto out;
1557         }
1558
1559         path = ocfs2_new_path_from_et(&et);
1560         if (!path) {
1561                 ret = -ENOMEM;
1562                 mlog_errno(ret);
1563                 goto out;
1564         }
1565
1566         while (trunc_end > trunc_start) {
1567
1568                 ret = ocfs2_find_path(INODE_CACHE(inode), path,
1569                                       cluster_in_el);
1570                 if (ret) {
1571                         mlog_errno(ret);
1572                         goto out;
1573                 }
1574
1575                 el = path_leaf_el(path);
1576
1577                 i = ocfs2_find_rec(el, trunc_end);
1578                 /*
1579                  * Need to go to previous extent block.
1580                  */
1581                 if (i < 0) {
1582                         if (path->p_tree_depth == 0)
1583                                 break;
1584
1585                         ret = ocfs2_find_cpos_for_left_leaf(inode->i_sb,
1586                                                             path,
1587                                                             &cluster_in_el);
1588                         if (ret) {
1589                                 mlog_errno(ret);
1590                                 goto out;
1591                         }
1592
1593                         /*
1594                          * We've reached the leftmost extent block,
1595                          * it's safe to leave.
1596                          */
1597                         if (cluster_in_el == 0)
1598                                 break;
1599
1600                         /*
1601                          * The 'pos' searched for previous extent block is
1602                          * always one cluster less than actual trunc_end.
1603                          */
1604                         trunc_end = cluster_in_el + 1;
1605
1606                         ocfs2_reinit_path(path, 1);
1607
1608                         continue;
1609
1610                 } else
1611                         rec = &el->l_recs[i];
1612
1613                 ocfs2_calc_trunc_pos(inode, el, rec, trunc_start, &trunc_cpos,
1614                                      &trunc_len, &trunc_end, &blkno, &done);
1615                 if (done)
1616                         break;
1617
1618                 flags = rec->e_flags;
1619                 phys_cpos = ocfs2_blocks_to_clusters(inode->i_sb, blkno);
1620
1621                 ret = ocfs2_remove_btree_range(inode, &et, trunc_cpos,
1622                                                phys_cpos, trunc_len, flags,
1623                                                &dealloc, refcount_loc);
1624                 if (ret < 0) {
1625                         mlog_errno(ret);
1626                         goto out;
1627                 }
1628
1629                 cluster_in_el = trunc_end;
1630
1631                 ocfs2_reinit_path(path, 1);
1632         }
1633
1634         ocfs2_truncate_cluster_pages(inode, byte_start, byte_len);
1635
1636 out:
1637         ocfs2_schedule_truncate_log_flush(osb, 1);
1638         ocfs2_run_deallocs(osb, &dealloc);
1639
1640         return ret;
1641 }
1642
1643 /*
1644  * Parts of this function taken from xfs_change_file_space()
1645  */
1646 static int __ocfs2_change_file_space(struct file *file, struct inode *inode,
1647                                      loff_t f_pos, unsigned int cmd,
1648                                      struct ocfs2_space_resv *sr,
1649                                      int change_size)
1650 {
1651         int ret;
1652         s64 llen;
1653         loff_t size;
1654         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1655         struct buffer_head *di_bh = NULL;
1656         handle_t *handle;
1657         unsigned long long max_off = inode->i_sb->s_maxbytes;
1658
1659         if (ocfs2_is_hard_readonly(osb) || ocfs2_is_soft_readonly(osb))
1660                 return -EROFS;
1661
1662         mutex_lock(&inode->i_mutex);
1663
1664         /*
1665          * This prevents concurrent writes on other nodes
1666          */
1667         ret = ocfs2_rw_lock(inode, 1);
1668         if (ret) {
1669                 mlog_errno(ret);
1670                 goto out;
1671         }
1672
1673         ret = ocfs2_inode_lock(inode, &di_bh, 1);
1674         if (ret) {
1675                 mlog_errno(ret);
1676                 goto out_rw_unlock;
1677         }
1678
1679         if (inode->i_flags & (S_IMMUTABLE|S_APPEND)) {
1680                 ret = -EPERM;
1681                 goto out_inode_unlock;
1682         }
1683
1684         switch (sr->l_whence) {
1685         case 0: /*SEEK_SET*/
1686                 break;
1687         case 1: /*SEEK_CUR*/
1688                 sr->l_start += f_pos;
1689                 break;
1690         case 2: /*SEEK_END*/
1691                 sr->l_start += i_size_read(inode);
1692                 break;
1693         default:
1694                 ret = -EINVAL;
1695                 goto out_inode_unlock;
1696         }
1697         sr->l_whence = 0;
1698
1699         llen = sr->l_len > 0 ? sr->l_len - 1 : sr->l_len;
1700
1701         if (sr->l_start < 0
1702             || sr->l_start > max_off
1703             || (sr->l_start + llen) < 0
1704             || (sr->l_start + llen) > max_off) {
1705                 ret = -EINVAL;
1706                 goto out_inode_unlock;
1707         }
1708         size = sr->l_start + sr->l_len;
1709
1710         if (cmd == OCFS2_IOC_RESVSP || cmd == OCFS2_IOC_RESVSP64) {
1711                 if (sr->l_len <= 0) {
1712                         ret = -EINVAL;
1713                         goto out_inode_unlock;
1714                 }
1715         }
1716
1717         if (file && should_remove_suid(file->f_path.dentry)) {
1718                 ret = __ocfs2_write_remove_suid(inode, di_bh);
1719                 if (ret) {
1720                         mlog_errno(ret);
1721                         goto out_inode_unlock;
1722                 }
1723         }
1724
1725         down_write(&OCFS2_I(inode)->ip_alloc_sem);
1726         switch (cmd) {
1727         case OCFS2_IOC_RESVSP:
1728         case OCFS2_IOC_RESVSP64:
1729                 /*
1730                  * This takes unsigned offsets, but the signed ones we
1731                  * pass have been checked against overflow above.
1732                  */
1733                 ret = ocfs2_allocate_unwritten_extents(inode, sr->l_start,
1734                                                        sr->l_len);
1735                 break;
1736         case OCFS2_IOC_UNRESVSP:
1737         case OCFS2_IOC_UNRESVSP64:
1738                 ret = ocfs2_remove_inode_range(inode, di_bh, sr->l_start,
1739                                                sr->l_len);
1740                 break;
1741         default:
1742                 ret = -EINVAL;
1743         }
1744         up_write(&OCFS2_I(inode)->ip_alloc_sem);
1745         if (ret) {
1746                 mlog_errno(ret);
1747                 goto out_inode_unlock;
1748         }
1749
1750         /*
1751          * We update c/mtime for these changes
1752          */
1753         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1754         if (IS_ERR(handle)) {
1755                 ret = PTR_ERR(handle);
1756                 mlog_errno(ret);
1757                 goto out_inode_unlock;
1758         }
1759
1760         if (change_size && i_size_read(inode) < size)
1761                 i_size_write(inode, size);
1762
1763         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
1764         ret = ocfs2_mark_inode_dirty(handle, inode, di_bh);
1765         if (ret < 0)
1766                 mlog_errno(ret);
1767
1768         ocfs2_commit_trans(osb, handle);
1769
1770 out_inode_unlock:
1771         brelse(di_bh);
1772         ocfs2_inode_unlock(inode, 1);
1773 out_rw_unlock:
1774         ocfs2_rw_unlock(inode, 1);
1775
1776 out:
1777         mutex_unlock(&inode->i_mutex);
1778         return ret;
1779 }
1780
1781 int ocfs2_change_file_space(struct file *file, unsigned int cmd,
1782                             struct ocfs2_space_resv *sr)
1783 {
1784         struct inode *inode = file->f_path.dentry->d_inode;
1785         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1786
1787         if ((cmd == OCFS2_IOC_RESVSP || cmd == OCFS2_IOC_RESVSP64) &&
1788             !ocfs2_writes_unwritten_extents(osb))
1789                 return -ENOTTY;
1790         else if ((cmd == OCFS2_IOC_UNRESVSP || cmd == OCFS2_IOC_UNRESVSP64) &&
1791                  !ocfs2_sparse_alloc(osb))
1792                 return -ENOTTY;
1793
1794         if (!S_ISREG(inode->i_mode))
1795                 return -EINVAL;
1796
1797         if (!(file->f_mode & FMODE_WRITE))
1798                 return -EBADF;
1799
1800         return __ocfs2_change_file_space(file, inode, file->f_pos, cmd, sr, 0);
1801 }
1802
1803 static long ocfs2_fallocate(struct inode *inode, int mode, loff_t offset,
1804                             loff_t len)
1805 {
1806         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1807         struct ocfs2_space_resv sr;
1808         int change_size = 1;
1809
1810         if (!ocfs2_writes_unwritten_extents(osb))
1811                 return -EOPNOTSUPP;
1812
1813         if (S_ISDIR(inode->i_mode))
1814                 return -ENODEV;
1815
1816         if (mode & FALLOC_FL_KEEP_SIZE)
1817                 change_size = 0;
1818
1819         sr.l_whence = 0;
1820         sr.l_start = (s64)offset;
1821         sr.l_len = (s64)len;
1822
1823         return __ocfs2_change_file_space(NULL, inode, offset,
1824                                          OCFS2_IOC_RESVSP64, &sr, change_size);
1825 }
1826
1827 int ocfs2_check_range_for_refcount(struct inode *inode, loff_t pos,
1828                                    size_t count)
1829 {
1830         int ret = 0;
1831         unsigned int extent_flags;
1832         u32 cpos, clusters, extent_len, phys_cpos;
1833         struct super_block *sb = inode->i_sb;
1834
1835         if (!ocfs2_refcount_tree(OCFS2_SB(inode->i_sb)) ||
1836             !(OCFS2_I(inode)->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL) ||
1837             OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
1838                 return 0;
1839
1840         cpos = pos >> OCFS2_SB(sb)->s_clustersize_bits;
1841         clusters = ocfs2_clusters_for_bytes(sb, pos + count) - cpos;
1842
1843         while (clusters) {
1844                 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos, &extent_len,
1845                                          &extent_flags);
1846                 if (ret < 0) {
1847                         mlog_errno(ret);
1848                         goto out;
1849                 }
1850
1851                 if (phys_cpos && (extent_flags & OCFS2_EXT_REFCOUNTED)) {
1852                         ret = 1;
1853                         break;
1854                 }
1855
1856                 if (extent_len > clusters)
1857                         extent_len = clusters;
1858
1859                 clusters -= extent_len;
1860                 cpos += extent_len;
1861         }
1862 out:
1863         return ret;
1864 }
1865
1866 static int ocfs2_prepare_inode_for_refcount(struct inode *inode,
1867                                             loff_t pos, size_t count,
1868                                             int *meta_level)
1869 {
1870         int ret;
1871         struct buffer_head *di_bh = NULL;
1872         u32 cpos = pos >> OCFS2_SB(inode->i_sb)->s_clustersize_bits;
1873         u32 clusters =
1874                 ocfs2_clusters_for_bytes(inode->i_sb, pos + count) - cpos;
1875
1876         ret = ocfs2_inode_lock(inode, &di_bh, 1);
1877         if (ret) {
1878                 mlog_errno(ret);
1879                 goto out;
1880         }
1881
1882         *meta_level = 1;
1883
1884         ret = ocfs2_refcount_cow(inode, di_bh, cpos, clusters, UINT_MAX);
1885         if (ret)
1886                 mlog_errno(ret);
1887 out:
1888         brelse(di_bh);
1889         return ret;
1890 }
1891
1892 static int ocfs2_prepare_inode_for_write(struct dentry *dentry,
1893                                          loff_t *ppos,
1894                                          size_t count,
1895                                          int appending,
1896                                          int *direct_io,
1897                                          int *has_refcount)
1898 {
1899         int ret = 0, meta_level = 0;
1900         struct inode *inode = dentry->d_inode;
1901         loff_t saved_pos, end;
1902
1903         /*
1904          * We start with a read level meta lock and only jump to an ex
1905          * if we need to make modifications here.
1906          */
1907         for(;;) {
1908                 ret = ocfs2_inode_lock(inode, NULL, meta_level);
1909                 if (ret < 0) {
1910                         meta_level = -1;
1911                         mlog_errno(ret);
1912                         goto out;
1913                 }
1914
1915                 /* Clear suid / sgid if necessary. We do this here
1916                  * instead of later in the write path because
1917                  * remove_suid() calls ->setattr without any hint that
1918                  * we may have already done our cluster locking. Since
1919                  * ocfs2_setattr() *must* take cluster locks to
1920                  * proceeed, this will lead us to recursively lock the
1921                  * inode. There's also the dinode i_size state which
1922                  * can be lost via setattr during extending writes (we
1923                  * set inode->i_size at the end of a write. */
1924                 if (should_remove_suid(dentry)) {
1925                         if (meta_level == 0) {
1926                                 ocfs2_inode_unlock(inode, meta_level);
1927                                 meta_level = 1;
1928                                 continue;
1929                         }
1930
1931                         ret = ocfs2_write_remove_suid(inode);
1932                         if (ret < 0) {
1933                                 mlog_errno(ret);
1934                                 goto out_unlock;
1935                         }
1936                 }
1937
1938                 /* work on a copy of ppos until we're sure that we won't have
1939                  * to recalculate it due to relocking. */
1940                 if (appending) {
1941                         saved_pos = i_size_read(inode);
1942                         mlog(0, "O_APPEND: inode->i_size=%llu\n", saved_pos);
1943                 } else {
1944                         saved_pos = *ppos;
1945                 }
1946
1947                 end = saved_pos + count;
1948
1949                 ret = ocfs2_check_range_for_refcount(inode, saved_pos, count);
1950                 if (ret == 1) {
1951                         ocfs2_inode_unlock(inode, meta_level);
1952                         meta_level = -1;
1953
1954                         ret = ocfs2_prepare_inode_for_refcount(inode,
1955                                                                saved_pos,
1956                                                                count,
1957                                                                &meta_level);
1958                         if (has_refcount)
1959                                 *has_refcount = 1;
1960                         if (direct_io)
1961                                 *direct_io = 0;
1962                 }
1963
1964                 if (ret < 0) {
1965                         mlog_errno(ret);
1966                         goto out_unlock;
1967                 }
1968
1969                 /*
1970                  * Skip the O_DIRECT checks if we don't need
1971                  * them.
1972                  */
1973                 if (!direct_io || !(*direct_io))
1974                         break;
1975
1976                 /*
1977                  * There's no sane way to do direct writes to an inode
1978                  * with inline data.
1979                  */
1980                 if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1981                         *direct_io = 0;
1982                         break;
1983                 }
1984
1985                 /*
1986                  * Allowing concurrent direct writes means
1987                  * i_size changes wouldn't be synchronized, so
1988                  * one node could wind up truncating another
1989                  * nodes writes.
1990                  */
1991                 if (end > i_size_read(inode)) {
1992                         *direct_io = 0;
1993                         break;
1994                 }
1995
1996                 /*
1997                  * We don't fill holes during direct io, so
1998                  * check for them here. If any are found, the
1999                  * caller will have to retake some cluster
2000                  * locks and initiate the io as buffered.
2001                  */
2002                 ret = ocfs2_check_range_for_holes(inode, saved_pos, count);
2003                 if (ret == 1) {
2004                         *direct_io = 0;
2005                         ret = 0;
2006                 } else if (ret < 0)
2007                         mlog_errno(ret);
2008                 break;
2009         }
2010
2011         if (appending)
2012                 *ppos = saved_pos;
2013
2014 out_unlock:
2015         if (meta_level >= 0)
2016                 ocfs2_inode_unlock(inode, meta_level);
2017
2018 out:
2019         return ret;
2020 }
2021
2022 static ssize_t ocfs2_file_aio_write(struct kiocb *iocb,
2023                                     const struct iovec *iov,
2024                                     unsigned long nr_segs,
2025                                     loff_t pos)
2026 {
2027         int ret, direct_io, appending, rw_level, have_alloc_sem  = 0;
2028         int can_do_direct, has_refcount = 0;
2029         ssize_t written = 0;
2030         size_t ocount;          /* original count */
2031         size_t count;           /* after file limit checks */
2032         loff_t old_size, *ppos = &iocb->ki_pos;
2033         u32 old_clusters;
2034         struct file *file = iocb->ki_filp;
2035         struct inode *inode = file->f_path.dentry->d_inode;
2036         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
2037
2038         mlog_entry("(0x%p, %u, '%.*s')\n", file,
2039                    (unsigned int)nr_segs,
2040                    file->f_path.dentry->d_name.len,
2041                    file->f_path.dentry->d_name.name);
2042
2043         if (iocb->ki_left == 0)
2044                 return 0;
2045
2046         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
2047
2048         appending = file->f_flags & O_APPEND ? 1 : 0;
2049         direct_io = file->f_flags & O_DIRECT ? 1 : 0;
2050
2051         mutex_lock(&inode->i_mutex);
2052
2053 relock:
2054         /* to match setattr's i_mutex -> i_alloc_sem -> rw_lock ordering */
2055         if (direct_io) {
2056                 down_read(&inode->i_alloc_sem);
2057                 have_alloc_sem = 1;
2058         }
2059
2060         /* concurrent O_DIRECT writes are allowed */
2061         rw_level = !direct_io;
2062         ret = ocfs2_rw_lock(inode, rw_level);
2063         if (ret < 0) {
2064                 mlog_errno(ret);
2065                 goto out_sems;
2066         }
2067
2068         can_do_direct = direct_io;
2069         ret = ocfs2_prepare_inode_for_write(file->f_path.dentry, ppos,
2070                                             iocb->ki_left, appending,
2071                                             &can_do_direct, &has_refcount);
2072         if (ret < 0) {
2073                 mlog_errno(ret);
2074                 goto out;
2075         }
2076
2077         /*
2078          * We can't complete the direct I/O as requested, fall back to
2079          * buffered I/O.
2080          */
2081         if (direct_io && !can_do_direct) {
2082                 ocfs2_rw_unlock(inode, rw_level);
2083                 up_read(&inode->i_alloc_sem);
2084
2085                 have_alloc_sem = 0;
2086                 rw_level = -1;
2087
2088                 direct_io = 0;
2089                 goto relock;
2090         }
2091
2092         /*
2093          * To later detect whether a journal commit for sync writes is
2094          * necessary, we sample i_size, and cluster count here.
2095          */
2096         old_size = i_size_read(inode);
2097         old_clusters = OCFS2_I(inode)->ip_clusters;
2098
2099         /* communicate with ocfs2_dio_end_io */
2100         ocfs2_iocb_set_rw_locked(iocb, rw_level);
2101
2102         ret = generic_segment_checks(iov, &nr_segs, &ocount,
2103                                      VERIFY_READ);
2104         if (ret)
2105                 goto out_dio;
2106
2107         count = ocount;
2108         ret = generic_write_checks(file, ppos, &count,
2109                                    S_ISBLK(inode->i_mode));
2110         if (ret)
2111                 goto out_dio;
2112
2113         if (direct_io) {
2114                 written = generic_file_direct_write(iocb, iov, &nr_segs, *ppos,
2115                                                     ppos, count, ocount);
2116                 if (written < 0) {
2117                         /*
2118                          * direct write may have instantiated a few
2119                          * blocks outside i_size. Trim these off again.
2120                          * Don't need i_size_read because we hold i_mutex.
2121                          *
2122                          * XXX(hch): this looks buggy because ocfs2 did not
2123                          * actually implement ->truncate.  Take a look at
2124                          * the new truncate sequence and update this accordingly
2125                          */
2126                         if (*ppos + count > inode->i_size)
2127                                 simple_setsize(inode, inode->i_size);
2128                         ret = written;
2129                         goto out_dio;
2130                 }
2131         } else {
2132                 current->backing_dev_info = file->f_mapping->backing_dev_info;
2133                 written = generic_file_buffered_write(iocb, iov, nr_segs, *ppos,
2134                                                       ppos, count, 0);
2135                 current->backing_dev_info = NULL;
2136         }
2137
2138 out_dio:
2139         /* buffered aio wouldn't have proper lock coverage today */
2140         BUG_ON(ret == -EIOCBQUEUED && !(file->f_flags & O_DIRECT));
2141
2142         if (((file->f_flags & O_DSYNC) && !direct_io) || IS_SYNC(inode) ||
2143             ((file->f_flags & O_DIRECT) && has_refcount)) {
2144                 ret = filemap_fdatawrite_range(file->f_mapping, pos,
2145                                                pos + count - 1);
2146                 if (ret < 0)
2147                         written = ret;
2148
2149                 if (!ret && ((old_size != i_size_read(inode)) ||
2150                              (old_clusters != OCFS2_I(inode)->ip_clusters) ||
2151                              has_refcount)) {
2152                         ret = jbd2_journal_force_commit(osb->journal->j_journal);
2153                         if (ret < 0)
2154                                 written = ret;
2155                 }
2156
2157                 if (!ret)
2158                         ret = filemap_fdatawait_range(file->f_mapping, pos,
2159                                                       pos + count - 1);
2160         }
2161
2162         /*
2163          * deep in g_f_a_w_n()->ocfs2_direct_IO we pass in a ocfs2_dio_end_io
2164          * function pointer which is called when o_direct io completes so that
2165          * it can unlock our rw lock.  (it's the clustered equivalent of
2166          * i_alloc_sem; protects truncate from racing with pending ios).
2167          * Unfortunately there are error cases which call end_io and others
2168          * that don't.  so we don't have to unlock the rw_lock if either an
2169          * async dio is going to do it in the future or an end_io after an
2170          * error has already done it.
2171          */
2172         if ((ret == -EIOCBQUEUED) || (!ocfs2_iocb_is_rw_locked(iocb))) {
2173                 rw_level = -1;
2174                 have_alloc_sem = 0;
2175         }
2176
2177 out:
2178         if (rw_level != -1)
2179                 ocfs2_rw_unlock(inode, rw_level);
2180
2181 out_sems:
2182         if (have_alloc_sem)
2183                 up_read(&inode->i_alloc_sem);
2184
2185         mutex_unlock(&inode->i_mutex);
2186
2187         if (written)
2188                 ret = written;
2189         mlog_exit(ret);
2190         return ret;
2191 }
2192
2193 static int ocfs2_splice_to_file(struct pipe_inode_info *pipe,
2194                                 struct file *out,
2195                                 struct splice_desc *sd)
2196 {
2197         int ret;
2198
2199         ret = ocfs2_prepare_inode_for_write(out->f_path.dentry, &sd->pos,
2200                                             sd->total_len, 0, NULL, NULL);
2201         if (ret < 0) {
2202                 mlog_errno(ret);
2203                 return ret;
2204         }
2205
2206         return splice_from_pipe_feed(pipe, sd, pipe_to_file);
2207 }
2208
2209 static ssize_t ocfs2_file_splice_write(struct pipe_inode_info *pipe,
2210                                        struct file *out,
2211                                        loff_t *ppos,
2212                                        size_t len,
2213                                        unsigned int flags)
2214 {
2215         int ret;
2216         struct address_space *mapping = out->f_mapping;
2217         struct inode *inode = mapping->host;
2218         struct splice_desc sd = {
2219                 .total_len = len,
2220                 .flags = flags,
2221                 .pos = *ppos,
2222                 .u.file = out,
2223         };
2224
2225         mlog_entry("(0x%p, 0x%p, %u, '%.*s')\n", out, pipe,
2226                    (unsigned int)len,
2227                    out->f_path.dentry->d_name.len,
2228                    out->f_path.dentry->d_name.name);
2229
2230         if (pipe->inode)
2231                 mutex_lock_nested(&pipe->inode->i_mutex, I_MUTEX_PARENT);
2232
2233         splice_from_pipe_begin(&sd);
2234         do {
2235                 ret = splice_from_pipe_next(pipe, &sd);
2236                 if (ret <= 0)
2237                         break;
2238
2239                 mutex_lock_nested(&inode->i_mutex, I_MUTEX_CHILD);
2240                 ret = ocfs2_rw_lock(inode, 1);
2241                 if (ret < 0)
2242                         mlog_errno(ret);
2243                 else {
2244                         ret = ocfs2_splice_to_file(pipe, out, &sd);
2245                         ocfs2_rw_unlock(inode, 1);
2246                 }
2247                 mutex_unlock(&inode->i_mutex);
2248         } while (ret > 0);
2249         splice_from_pipe_end(pipe, &sd);
2250
2251         if (pipe->inode)
2252                 mutex_unlock(&pipe->inode->i_mutex);
2253
2254         if (sd.num_spliced)
2255                 ret = sd.num_spliced;
2256
2257         if (ret > 0) {
2258                 unsigned long nr_pages;
2259                 int err;
2260
2261                 nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
2262
2263                 err = generic_write_sync(out, *ppos, ret);
2264                 if (err)
2265                         ret = err;
2266                 else
2267                         *ppos += ret;
2268
2269                 balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
2270         }
2271
2272         mlog_exit(ret);
2273         return ret;
2274 }
2275
2276 static ssize_t ocfs2_file_splice_read(struct file *in,
2277                                       loff_t *ppos,
2278                                       struct pipe_inode_info *pipe,
2279                                       size_t len,
2280                                       unsigned int flags)
2281 {
2282         int ret = 0, lock_level = 0;
2283         struct inode *inode = in->f_path.dentry->d_inode;
2284
2285         mlog_entry("(0x%p, 0x%p, %u, '%.*s')\n", in, pipe,
2286                    (unsigned int)len,
2287                    in->f_path.dentry->d_name.len,
2288                    in->f_path.dentry->d_name.name);
2289
2290         /*
2291          * See the comment in ocfs2_file_aio_read()
2292          */
2293         ret = ocfs2_inode_lock_atime(inode, in->f_vfsmnt, &lock_level);
2294         if (ret < 0) {
2295                 mlog_errno(ret);
2296                 goto bail;
2297         }
2298         ocfs2_inode_unlock(inode, lock_level);
2299
2300         ret = generic_file_splice_read(in, ppos, pipe, len, flags);
2301
2302 bail:
2303         mlog_exit(ret);
2304         return ret;
2305 }
2306
2307 static ssize_t ocfs2_file_aio_read(struct kiocb *iocb,
2308                                    const struct iovec *iov,
2309                                    unsigned long nr_segs,
2310                                    loff_t pos)
2311 {
2312         int ret = 0, rw_level = -1, have_alloc_sem = 0, lock_level = 0;
2313         struct file *filp = iocb->ki_filp;
2314         struct inode *inode = filp->f_path.dentry->d_inode;
2315
2316         mlog_entry("(0x%p, %u, '%.*s')\n", filp,
2317                    (unsigned int)nr_segs,
2318                    filp->f_path.dentry->d_name.len,
2319                    filp->f_path.dentry->d_name.name);
2320
2321         if (!inode) {
2322                 ret = -EINVAL;
2323                 mlog_errno(ret);
2324                 goto bail;
2325         }
2326
2327         /*
2328          * buffered reads protect themselves in ->readpage().  O_DIRECT reads
2329          * need locks to protect pending reads from racing with truncate.
2330          */
2331         if (filp->f_flags & O_DIRECT) {
2332                 down_read(&inode->i_alloc_sem);
2333                 have_alloc_sem = 1;
2334
2335                 ret = ocfs2_rw_lock(inode, 0);
2336                 if (ret < 0) {
2337                         mlog_errno(ret);
2338                         goto bail;
2339                 }
2340                 rw_level = 0;
2341                 /* communicate with ocfs2_dio_end_io */
2342                 ocfs2_iocb_set_rw_locked(iocb, rw_level);
2343         }
2344
2345         /*
2346          * We're fine letting folks race truncates and extending
2347          * writes with read across the cluster, just like they can
2348          * locally. Hence no rw_lock during read.
2349          *
2350          * Take and drop the meta data lock to update inode fields
2351          * like i_size. This allows the checks down below
2352          * generic_file_aio_read() a chance of actually working.
2353          */
2354         ret = ocfs2_inode_lock_atime(inode, filp->f_vfsmnt, &lock_level);
2355         if (ret < 0) {
2356                 mlog_errno(ret);
2357                 goto bail;
2358         }
2359         ocfs2_inode_unlock(inode, lock_level);
2360
2361         ret = generic_file_aio_read(iocb, iov, nr_segs, iocb->ki_pos);
2362         if (ret == -EINVAL)
2363                 mlog(0, "generic_file_aio_read returned -EINVAL\n");
2364
2365         /* buffered aio wouldn't have proper lock coverage today */
2366         BUG_ON(ret == -EIOCBQUEUED && !(filp->f_flags & O_DIRECT));
2367
2368         /* see ocfs2_file_aio_write */
2369         if (ret == -EIOCBQUEUED || !ocfs2_iocb_is_rw_locked(iocb)) {
2370                 rw_level = -1;
2371                 have_alloc_sem = 0;
2372         }
2373
2374 bail:
2375         if (have_alloc_sem)
2376                 up_read(&inode->i_alloc_sem);
2377         if (rw_level != -1)
2378                 ocfs2_rw_unlock(inode, rw_level);
2379         mlog_exit(ret);
2380
2381         return ret;
2382 }
2383
2384 const struct inode_operations ocfs2_file_iops = {
2385         .setattr        = ocfs2_setattr,
2386         .getattr        = ocfs2_getattr,
2387         .permission     = ocfs2_permission,
2388         .setxattr       = generic_setxattr,
2389         .getxattr       = generic_getxattr,
2390         .listxattr      = ocfs2_listxattr,
2391         .removexattr    = generic_removexattr,
2392         .fallocate      = ocfs2_fallocate,
2393         .fiemap         = ocfs2_fiemap,
2394 };
2395
2396 const struct inode_operations ocfs2_special_file_iops = {
2397         .setattr        = ocfs2_setattr,
2398         .getattr        = ocfs2_getattr,
2399         .permission     = ocfs2_permission,
2400 };
2401
2402 /*
2403  * Other than ->lock, keep ocfs2_fops and ocfs2_dops in sync with
2404  * ocfs2_fops_no_plocks and ocfs2_dops_no_plocks!
2405  */
2406 const struct file_operations ocfs2_fops = {
2407         .llseek         = generic_file_llseek,
2408         .read           = do_sync_read,
2409         .write          = do_sync_write,
2410         .mmap           = ocfs2_mmap,
2411         .fsync          = ocfs2_sync_file,
2412         .release        = ocfs2_file_release,
2413         .open           = ocfs2_file_open,
2414         .aio_read       = ocfs2_file_aio_read,
2415         .aio_write      = ocfs2_file_aio_write,
2416         .unlocked_ioctl = ocfs2_ioctl,
2417 #ifdef CONFIG_COMPAT
2418         .compat_ioctl   = ocfs2_compat_ioctl,
2419 #endif
2420         .lock           = ocfs2_lock,
2421         .flock          = ocfs2_flock,
2422         .splice_read    = ocfs2_file_splice_read,
2423         .splice_write   = ocfs2_file_splice_write,
2424 };
2425
2426 const struct file_operations ocfs2_dops = {
2427         .llseek         = generic_file_llseek,
2428         .read           = generic_read_dir,
2429         .readdir        = ocfs2_readdir,
2430         .fsync          = ocfs2_sync_file,
2431         .release        = ocfs2_dir_release,
2432         .open           = ocfs2_dir_open,
2433         .unlocked_ioctl = ocfs2_ioctl,
2434 #ifdef CONFIG_COMPAT
2435         .compat_ioctl   = ocfs2_compat_ioctl,
2436 #endif
2437         .lock           = ocfs2_lock,
2438         .flock          = ocfs2_flock,
2439 };
2440
2441 /*
2442  * POSIX-lockless variants of our file_operations.
2443  *
2444  * These will be used if the underlying cluster stack does not support
2445  * posix file locking, if the user passes the "localflocks" mount
2446  * option, or if we have a local-only fs.
2447  *
2448  * ocfs2_flock is in here because all stacks handle UNIX file locks,
2449  * so we still want it in the case of no stack support for
2450  * plocks. Internally, it will do the right thing when asked to ignore
2451  * the cluster.
2452  */
2453 const struct file_operations ocfs2_fops_no_plocks = {
2454         .llseek         = generic_file_llseek,
2455         .read           = do_sync_read,
2456         .write          = do_sync_write,
2457         .mmap           = ocfs2_mmap,
2458         .fsync          = ocfs2_sync_file,
2459         .release        = ocfs2_file_release,
2460         .open           = ocfs2_file_open,
2461         .aio_read       = ocfs2_file_aio_read,
2462         .aio_write      = ocfs2_file_aio_write,
2463         .unlocked_ioctl = ocfs2_ioctl,
2464 #ifdef CONFIG_COMPAT
2465         .compat_ioctl   = ocfs2_compat_ioctl,
2466 #endif
2467         .flock          = ocfs2_flock,
2468         .splice_read    = ocfs2_file_splice_read,
2469         .splice_write   = ocfs2_file_splice_write,
2470 };
2471
2472 const struct file_operations ocfs2_dops_no_plocks = {
2473         .llseek         = generic_file_llseek,
2474         .read           = generic_read_dir,
2475         .readdir        = ocfs2_readdir,
2476         .fsync          = ocfs2_sync_file,
2477         .release        = ocfs2_dir_release,
2478         .open           = ocfs2_dir_open,
2479         .unlocked_ioctl = ocfs2_ioctl,
2480 #ifdef CONFIG_COMPAT
2481         .compat_ioctl   = ocfs2_compat_ioctl,
2482 #endif
2483         .flock          = ocfs2_flock,
2484 };