1 /* -*- mode: c; c-basic-offset: 8; -*-
2 * vim: noexpandtab sw=8 ts=8 sts=0:
6 * File open, close, extend, truncate
8 * Copyright (C) 2002, 2004 Oracle. All rights reserved.
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.
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.
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.
26 #include <linux/capability.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>
40 #define MLOG_MASK_PREFIX ML_INODE
41 #include <cluster/masklog.h>
49 #include "extent_map.h"
62 #include "refcounttree.h"
64 #include "buffer_head_io.h"
66 static int ocfs2_sync_inode(struct inode *inode)
68 filemap_fdatawrite(inode->i_mapping);
69 return sync_mapping_buffers(inode->i_mapping);
72 static int ocfs2_init_file_private(struct inode *inode, struct file *file)
74 struct ocfs2_file_private *fp;
76 fp = kzalloc(sizeof(struct ocfs2_file_private), GFP_KERNEL);
81 mutex_init(&fp->fp_mutex);
82 ocfs2_file_lock_res_init(&fp->fp_flock, fp);
83 file->private_data = fp;
88 static void ocfs2_free_file_private(struct inode *inode, struct file *file)
90 struct ocfs2_file_private *fp = file->private_data;
91 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
94 ocfs2_simple_drop_lockres(osb, &fp->fp_flock);
95 ocfs2_lock_res_free(&fp->fp_flock);
97 file->private_data = NULL;
101 static int ocfs2_file_open(struct inode *inode, struct file *file)
104 int mode = file->f_flags;
105 struct ocfs2_inode_info *oi = OCFS2_I(inode);
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);
110 if (file->f_mode & FMODE_WRITE)
111 dquot_initialize(inode);
113 spin_lock(&oi->ip_lock);
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);
126 oi->ip_flags |= OCFS2_INODE_OPEN_DIRECT;
129 spin_unlock(&oi->ip_lock);
131 status = ocfs2_init_file_private(inode, file);
134 * We want to set open count back if we're failing the
137 spin_lock(&oi->ip_lock);
139 spin_unlock(&oi->ip_lock);
147 static int ocfs2_file_release(struct inode *inode, struct file *file)
149 struct ocfs2_inode_info *oi = OCFS2_I(inode);
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);
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);
160 ocfs2_free_file_private(inode, file);
167 static int ocfs2_dir_open(struct inode *inode, struct file *file)
169 return ocfs2_init_file_private(inode, file);
172 static int ocfs2_dir_release(struct inode *inode, struct file *file)
174 ocfs2_free_file_private(inode, file);
178 static int ocfs2_sync_file(struct file *file,
179 struct dentry *dentry,
184 struct inode *inode = dentry->d_inode;
185 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
187 mlog_entry("(0x%p, 0x%p, %d, '%.*s')\n", file, dentry, datasync,
188 dentry->d_name.len, dentry->d_name.name);
190 err = ocfs2_sync_inode(dentry->d_inode);
194 if (datasync && !(inode->i_state & I_DIRTY_DATASYNC))
197 journal = osb->journal->j_journal;
198 err = jbd2_journal_force_commit(journal);
203 return (err < 0) ? -EIO : 0;
206 int ocfs2_should_update_atime(struct inode *inode,
207 struct vfsmount *vfsmnt)
210 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
212 if (ocfs2_is_hard_readonly(osb) || ocfs2_is_soft_readonly(osb))
215 if ((inode->i_flags & S_NOATIME) ||
216 ((inode->i_sb->s_flags & MS_NODIRATIME) && S_ISDIR(inode->i_mode)))
220 * We can be called with no vfsmnt structure - NFSD will
223 * Note that our action here is different than touch_atime() -
224 * if we can't tell whether this is a noatime mount, then we
225 * don't know whether to trust the value of s_atime_quantum.
230 if ((vfsmnt->mnt_flags & MNT_NOATIME) ||
231 ((vfsmnt->mnt_flags & MNT_NODIRATIME) && S_ISDIR(inode->i_mode)))
234 if (vfsmnt->mnt_flags & MNT_RELATIME) {
235 if ((timespec_compare(&inode->i_atime, &inode->i_mtime) <= 0) ||
236 (timespec_compare(&inode->i_atime, &inode->i_ctime) <= 0))
243 if ((now.tv_sec - inode->i_atime.tv_sec <= osb->s_atime_quantum))
249 int ocfs2_update_inode_atime(struct inode *inode,
250 struct buffer_head *bh)
253 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
255 struct ocfs2_dinode *di = (struct ocfs2_dinode *) bh->b_data;
259 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
260 if (IS_ERR(handle)) {
261 ret = PTR_ERR(handle);
266 ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), bh,
267 OCFS2_JOURNAL_ACCESS_WRITE);
274 * Don't use ocfs2_mark_inode_dirty() here as we don't always
275 * have i_mutex to guard against concurrent changes to other
278 inode->i_atime = CURRENT_TIME;
279 di->i_atime = cpu_to_le64(inode->i_atime.tv_sec);
280 di->i_atime_nsec = cpu_to_le32(inode->i_atime.tv_nsec);
282 ret = ocfs2_journal_dirty(handle, bh);
287 ocfs2_commit_trans(OCFS2_SB(inode->i_sb), handle);
293 static int ocfs2_set_inode_size(handle_t *handle,
295 struct buffer_head *fe_bh,
301 i_size_write(inode, new_i_size);
302 inode->i_blocks = ocfs2_inode_sector_count(inode);
303 inode->i_ctime = inode->i_mtime = CURRENT_TIME;
305 status = ocfs2_mark_inode_dirty(handle, inode, fe_bh);
316 int ocfs2_simple_size_update(struct inode *inode,
317 struct buffer_head *di_bh,
321 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
322 handle_t *handle = NULL;
324 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
325 if (IS_ERR(handle)) {
326 ret = PTR_ERR(handle);
331 ret = ocfs2_set_inode_size(handle, inode, di_bh,
336 ocfs2_commit_trans(osb, handle);
341 static int ocfs2_cow_file_pos(struct inode *inode,
342 struct buffer_head *fe_bh,
346 u32 phys, cpos = offset >> OCFS2_SB(inode->i_sb)->s_clustersize_bits;
347 unsigned int num_clusters = 0;
348 unsigned int ext_flags = 0;
351 * If the new offset is aligned to the range of the cluster, there is
352 * no space for ocfs2_zero_range_for_truncate to fill, so no need to
355 if ((offset & (OCFS2_SB(inode->i_sb)->s_clustersize - 1)) == 0)
358 status = ocfs2_get_clusters(inode, cpos, &phys,
359 &num_clusters, &ext_flags);
365 if (!(ext_flags & OCFS2_EXT_REFCOUNTED))
368 return ocfs2_refcount_cow(inode, fe_bh, cpos, 1, cpos+1);
374 static int ocfs2_orphan_for_truncate(struct ocfs2_super *osb,
376 struct buffer_head *fe_bh,
381 struct ocfs2_dinode *di;
387 * We need to CoW the cluster contains the offset if it is reflinked
388 * since we will call ocfs2_zero_range_for_truncate later which will
389 * write "0" from offset to the end of the cluster.
391 status = ocfs2_cow_file_pos(inode, fe_bh, new_i_size);
397 /* TODO: This needs to actually orphan the inode in this
400 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
401 if (IS_ERR(handle)) {
402 status = PTR_ERR(handle);
407 status = ocfs2_journal_access_di(handle, INODE_CACHE(inode), fe_bh,
408 OCFS2_JOURNAL_ACCESS_WRITE);
415 * Do this before setting i_size.
417 cluster_bytes = ocfs2_align_bytes_to_clusters(inode->i_sb, new_i_size);
418 status = ocfs2_zero_range_for_truncate(inode, handle, new_i_size,
425 i_size_write(inode, new_i_size);
426 inode->i_ctime = inode->i_mtime = CURRENT_TIME;
428 di = (struct ocfs2_dinode *) fe_bh->b_data;
429 di->i_size = cpu_to_le64(new_i_size);
430 di->i_ctime = di->i_mtime = cpu_to_le64(inode->i_ctime.tv_sec);
431 di->i_ctime_nsec = di->i_mtime_nsec = cpu_to_le32(inode->i_ctime.tv_nsec);
433 status = ocfs2_journal_dirty(handle, fe_bh);
438 ocfs2_commit_trans(osb, handle);
445 static int ocfs2_truncate_file(struct inode *inode,
446 struct buffer_head *di_bh,
450 struct ocfs2_dinode *fe = NULL;
451 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
452 struct ocfs2_truncate_context *tc = NULL;
454 mlog_entry("(inode = %llu, new_i_size = %llu\n",
455 (unsigned long long)OCFS2_I(inode)->ip_blkno,
456 (unsigned long long)new_i_size);
458 /* We trust di_bh because it comes from ocfs2_inode_lock(), which
459 * already validated it */
460 fe = (struct ocfs2_dinode *) di_bh->b_data;
462 mlog_bug_on_msg(le64_to_cpu(fe->i_size) != i_size_read(inode),
463 "Inode %llu, inode i_size = %lld != di "
464 "i_size = %llu, i_flags = 0x%x\n",
465 (unsigned long long)OCFS2_I(inode)->ip_blkno,
467 (unsigned long long)le64_to_cpu(fe->i_size),
468 le32_to_cpu(fe->i_flags));
470 if (new_i_size > le64_to_cpu(fe->i_size)) {
471 mlog(0, "asked to truncate file with size (%llu) to size (%llu)!\n",
472 (unsigned long long)le64_to_cpu(fe->i_size),
473 (unsigned long long)new_i_size);
479 mlog(0, "inode %llu, i_size = %llu, new_i_size = %llu\n",
480 (unsigned long long)le64_to_cpu(fe->i_blkno),
481 (unsigned long long)le64_to_cpu(fe->i_size),
482 (unsigned long long)new_i_size);
484 /* lets handle the simple truncate cases before doing any more
485 * cluster locking. */
486 if (new_i_size == le64_to_cpu(fe->i_size))
489 down_write(&OCFS2_I(inode)->ip_alloc_sem);
492 * The inode lock forced other nodes to sync and drop their
493 * pages, which (correctly) happens even if we have a truncate
494 * without allocation change - ocfs2 cluster sizes can be much
495 * greater than page size, so we have to truncate them
498 unmap_mapping_range(inode->i_mapping, new_i_size + PAGE_SIZE - 1, 0, 1);
499 truncate_inode_pages(inode->i_mapping, new_i_size);
501 if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
502 status = ocfs2_truncate_inline(inode, di_bh, new_i_size,
503 i_size_read(inode), 1);
507 goto bail_unlock_sem;
510 /* alright, we're going to need to do a full blown alloc size
511 * change. Orphan the inode so that recovery can complete the
512 * truncate if necessary. This does the task of marking
514 status = ocfs2_orphan_for_truncate(osb, inode, di_bh, new_i_size);
517 goto bail_unlock_sem;
520 status = ocfs2_prepare_truncate(osb, inode, di_bh, &tc);
523 goto bail_unlock_sem;
526 status = ocfs2_commit_truncate(osb, inode, di_bh, tc);
529 goto bail_unlock_sem;
532 /* TODO: orphan dir cleanup here. */
534 up_write(&OCFS2_I(inode)->ip_alloc_sem);
537 if (!status && OCFS2_I(inode)->ip_clusters == 0)
538 status = ocfs2_try_remove_refcount_tree(inode, di_bh);
545 * extend file allocation only here.
546 * we'll update all the disk stuff, and oip->alloc_size
548 * expect stuff to be locked, a transaction started and enough data /
549 * metadata reservations in the contexts.
551 * Will return -EAGAIN, and a reason if a restart is needed.
552 * If passed in, *reason will always be set, even in error.
554 int ocfs2_add_inode_data(struct ocfs2_super *osb,
559 struct buffer_head *fe_bh,
561 struct ocfs2_alloc_context *data_ac,
562 struct ocfs2_alloc_context *meta_ac,
563 enum ocfs2_alloc_restarted *reason_ret)
566 struct ocfs2_extent_tree et;
568 ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), fe_bh);
569 ret = ocfs2_add_clusters_in_btree(handle, &et, logical_offset,
570 clusters_to_add, mark_unwritten,
571 data_ac, meta_ac, reason_ret);
576 static int __ocfs2_extend_allocation(struct inode *inode, u32 logical_start,
577 u32 clusters_to_add, int mark_unwritten)
580 int restart_func = 0;
583 struct buffer_head *bh = NULL;
584 struct ocfs2_dinode *fe = NULL;
585 handle_t *handle = NULL;
586 struct ocfs2_alloc_context *data_ac = NULL;
587 struct ocfs2_alloc_context *meta_ac = NULL;
588 enum ocfs2_alloc_restarted why;
589 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
590 struct ocfs2_extent_tree et;
593 mlog_entry("(clusters_to_add = %u)\n", clusters_to_add);
596 * This function only exists for file systems which don't
599 BUG_ON(mark_unwritten && !ocfs2_sparse_alloc(osb));
601 status = ocfs2_read_inode_block(inode, &bh);
606 fe = (struct ocfs2_dinode *) bh->b_data;
609 BUG_ON(le32_to_cpu(fe->i_clusters) != OCFS2_I(inode)->ip_clusters);
611 mlog(0, "extend inode %llu, i_size = %lld, di->i_clusters = %u, "
612 "clusters_to_add = %u\n",
613 (unsigned long long)OCFS2_I(inode)->ip_blkno,
614 (long long)i_size_read(inode), le32_to_cpu(fe->i_clusters),
616 ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), bh);
617 status = ocfs2_lock_allocators(inode, &et, clusters_to_add, 0,
624 credits = ocfs2_calc_extend_credits(osb->sb, &fe->id2.i_list,
626 handle = ocfs2_start_trans(osb, credits);
627 if (IS_ERR(handle)) {
628 status = PTR_ERR(handle);
634 restarted_transaction:
635 status = dquot_alloc_space_nodirty(inode,
636 ocfs2_clusters_to_bytes(osb->sb, clusters_to_add));
641 /* reserve a write to the file entry early on - that we if we
642 * run out of credits in the allocation path, we can still
644 status = ocfs2_journal_access_di(handle, INODE_CACHE(inode), bh,
645 OCFS2_JOURNAL_ACCESS_WRITE);
651 prev_clusters = OCFS2_I(inode)->ip_clusters;
653 status = ocfs2_add_inode_data(osb,
663 if ((status < 0) && (status != -EAGAIN)) {
664 if (status != -ENOSPC)
669 status = ocfs2_journal_dirty(handle, bh);
675 spin_lock(&OCFS2_I(inode)->ip_lock);
676 clusters_to_add -= (OCFS2_I(inode)->ip_clusters - prev_clusters);
677 spin_unlock(&OCFS2_I(inode)->ip_lock);
678 /* Release unused quota reservation */
679 dquot_free_space(inode,
680 ocfs2_clusters_to_bytes(osb->sb, clusters_to_add));
683 if (why != RESTART_NONE && clusters_to_add) {
684 if (why == RESTART_META) {
685 mlog(0, "restarting function.\n");
688 BUG_ON(why != RESTART_TRANS);
690 mlog(0, "restarting transaction.\n");
691 /* TODO: This can be more intelligent. */
692 credits = ocfs2_calc_extend_credits(osb->sb,
695 status = ocfs2_extend_trans(handle, credits);
697 /* handle still has to be committed at
703 goto restarted_transaction;
707 mlog(0, "fe: i_clusters = %u, i_size=%llu\n",
708 le32_to_cpu(fe->i_clusters),
709 (unsigned long long)le64_to_cpu(fe->i_size));
710 mlog(0, "inode: ip_clusters=%u, i_size=%lld\n",
711 OCFS2_I(inode)->ip_clusters, (long long)i_size_read(inode));
714 if (status < 0 && did_quota)
715 dquot_free_space(inode,
716 ocfs2_clusters_to_bytes(osb->sb, clusters_to_add));
718 ocfs2_commit_trans(osb, handle);
722 ocfs2_free_alloc_context(data_ac);
726 ocfs2_free_alloc_context(meta_ac);
729 if ((!status) && restart_func) {
740 /* Some parts of this taken from generic_cont_expand, which turned out
741 * to be too fragile to do exactly what we need without us having to
742 * worry about recursive locking in ->write_begin() and ->write_end(). */
743 static int ocfs2_write_zero_page(struct inode *inode,
746 struct address_space *mapping = inode->i_mapping;
750 handle_t *handle = NULL;
753 offset = (size & (PAGE_CACHE_SIZE-1)); /* Within page */
754 /* ugh. in prepare/commit_write, if from==to==start of block, we
755 ** skip the prepare. make sure we never send an offset for the start
758 if ((offset & (inode->i_sb->s_blocksize - 1)) == 0) {
761 index = size >> PAGE_CACHE_SHIFT;
763 page = grab_cache_page(mapping, index);
770 ret = ocfs2_prepare_write_nolock(inode, page, offset, offset);
776 if (ocfs2_should_order_data(inode)) {
777 handle = ocfs2_start_walk_page_trans(inode, page, offset,
779 if (IS_ERR(handle)) {
780 ret = PTR_ERR(handle);
786 /* must not update i_size! */
787 ret = block_commit_write(page, offset, offset);
794 ocfs2_commit_trans(OCFS2_SB(inode->i_sb), handle);
797 page_cache_release(page);
802 static int ocfs2_zero_extend(struct inode *inode,
807 struct super_block *sb = inode->i_sb;
809 start_off = ocfs2_align_bytes_to_blocks(sb, i_size_read(inode));
810 while (start_off < zero_to_size) {
811 ret = ocfs2_write_zero_page(inode, start_off);
817 start_off += sb->s_blocksize;
820 * Very large extends have the potential to lock up
821 * the cpu for extended periods of time.
830 int ocfs2_extend_no_holes(struct inode *inode, u64 new_i_size, u64 zero_to)
834 struct ocfs2_inode_info *oi = OCFS2_I(inode);
836 clusters_to_add = ocfs2_clusters_for_bytes(inode->i_sb, new_i_size);
837 if (clusters_to_add < oi->ip_clusters)
840 clusters_to_add -= oi->ip_clusters;
842 if (clusters_to_add) {
843 ret = __ocfs2_extend_allocation(inode, oi->ip_clusters,
852 * Call this even if we don't add any clusters to the tree. We
853 * still need to zero the area between the old i_size and the
856 ret = ocfs2_zero_extend(inode, zero_to);
864 static int ocfs2_extend_file(struct inode *inode,
865 struct buffer_head *di_bh,
869 struct ocfs2_inode_info *oi = OCFS2_I(inode);
873 /* setattr sometimes calls us like this. */
877 if (i_size_read(inode) == new_i_size)
879 BUG_ON(new_i_size < i_size_read(inode));
882 * Fall through for converting inline data, even if the fs
883 * supports sparse files.
885 * The check for inline data here is legal - nobody can add
886 * the feature since we have i_mutex. We must check it again
887 * after acquiring ip_alloc_sem though, as paths like mmap
888 * might have raced us to converting the inode to extents.
890 if (!(oi->ip_dyn_features & OCFS2_INLINE_DATA_FL)
891 && ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)))
892 goto out_update_size;
895 * The alloc sem blocks people in read/write from reading our
896 * allocation until we're done changing it. We depend on
897 * i_mutex to block other extend/truncate calls while we're
900 down_write(&oi->ip_alloc_sem);
902 if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
904 * We can optimize small extends by keeping the inodes
907 if (ocfs2_size_fits_inline_data(di_bh, new_i_size)) {
908 up_write(&oi->ip_alloc_sem);
909 goto out_update_size;
912 ret = ocfs2_convert_inline_data_to_extents(inode, di_bh);
914 up_write(&oi->ip_alloc_sem);
921 if (!ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)))
922 ret = ocfs2_extend_no_holes(inode, new_i_size, new_i_size);
924 up_write(&oi->ip_alloc_sem);
932 ret = ocfs2_simple_size_update(inode, di_bh, new_i_size);
940 int ocfs2_setattr(struct dentry *dentry, struct iattr *attr)
942 int status = 0, size_change;
943 struct inode *inode = dentry->d_inode;
944 struct super_block *sb = inode->i_sb;
945 struct ocfs2_super *osb = OCFS2_SB(sb);
946 struct buffer_head *bh = NULL;
947 handle_t *handle = NULL;
949 struct dquot *transfer_from[MAXQUOTAS] = { };
950 struct dquot *transfer_to[MAXQUOTAS] = { };
952 mlog_entry("(0x%p, '%.*s')\n", dentry,
953 dentry->d_name.len, dentry->d_name.name);
955 /* ensuring we don't even attempt to truncate a symlink */
956 if (S_ISLNK(inode->i_mode))
957 attr->ia_valid &= ~ATTR_SIZE;
959 if (attr->ia_valid & ATTR_MODE)
960 mlog(0, "mode change: %d\n", attr->ia_mode);
961 if (attr->ia_valid & ATTR_UID)
962 mlog(0, "uid change: %d\n", attr->ia_uid);
963 if (attr->ia_valid & ATTR_GID)
964 mlog(0, "gid change: %d\n", attr->ia_gid);
965 if (attr->ia_valid & ATTR_SIZE)
966 mlog(0, "size change...\n");
967 if (attr->ia_valid & (ATTR_ATIME | ATTR_MTIME | ATTR_CTIME))
968 mlog(0, "time change...\n");
970 #define OCFS2_VALID_ATTRS (ATTR_ATIME | ATTR_MTIME | ATTR_CTIME | ATTR_SIZE \
971 | ATTR_GID | ATTR_UID | ATTR_MODE)
972 if (!(attr->ia_valid & OCFS2_VALID_ATTRS)) {
973 mlog(0, "can't handle attrs: 0x%x\n", attr->ia_valid);
977 status = inode_change_ok(inode, attr);
981 size_change = S_ISREG(inode->i_mode) && attr->ia_valid & ATTR_SIZE;
983 dquot_initialize(inode);
985 status = ocfs2_rw_lock(inode, 1);
992 status = ocfs2_inode_lock(inode, &bh, 1);
994 if (status != -ENOENT)
999 if (size_change && attr->ia_size != i_size_read(inode)) {
1000 status = inode_newsize_ok(inode, attr->ia_size);
1004 if (i_size_read(inode) > attr->ia_size) {
1005 if (ocfs2_should_order_data(inode)) {
1006 status = ocfs2_begin_ordered_truncate(inode,
1011 status = ocfs2_truncate_file(inode, bh, attr->ia_size);
1013 status = ocfs2_extend_file(inode, bh, attr->ia_size);
1015 if (status != -ENOSPC)
1022 if ((attr->ia_valid & ATTR_UID && attr->ia_uid != inode->i_uid) ||
1023 (attr->ia_valid & ATTR_GID && attr->ia_gid != inode->i_gid)) {
1025 * Gather pointers to quota structures so that allocation /
1026 * freeing of quota structures happens here and not inside
1027 * dquot_transfer() where we have problems with lock ordering
1029 if (attr->ia_valid & ATTR_UID && attr->ia_uid != inode->i_uid
1030 && OCFS2_HAS_RO_COMPAT_FEATURE(sb,
1031 OCFS2_FEATURE_RO_COMPAT_USRQUOTA)) {
1032 transfer_to[USRQUOTA] = dqget(sb, attr->ia_uid,
1034 transfer_from[USRQUOTA] = dqget(sb, inode->i_uid,
1036 if (!transfer_to[USRQUOTA] || !transfer_from[USRQUOTA]) {
1041 if (attr->ia_valid & ATTR_GID && attr->ia_gid != inode->i_gid
1042 && OCFS2_HAS_RO_COMPAT_FEATURE(sb,
1043 OCFS2_FEATURE_RO_COMPAT_GRPQUOTA)) {
1044 transfer_to[GRPQUOTA] = dqget(sb, attr->ia_gid,
1046 transfer_from[GRPQUOTA] = dqget(sb, inode->i_gid,
1048 if (!transfer_to[GRPQUOTA] || !transfer_from[GRPQUOTA]) {
1053 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS +
1054 2 * ocfs2_quota_trans_credits(sb));
1055 if (IS_ERR(handle)) {
1056 status = PTR_ERR(handle);
1060 status = dquot_transfer(inode, attr);
1064 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1065 if (IS_ERR(handle)) {
1066 status = PTR_ERR(handle);
1073 * This will intentionally not wind up calling vmtruncate(),
1074 * since all the work for a size change has been done above.
1075 * Otherwise, we could get into problems with truncate as
1076 * ip_alloc_sem is used there to protect against i_size
1079 status = inode_setattr(inode, attr);
1085 status = ocfs2_mark_inode_dirty(handle, inode, bh);
1090 ocfs2_commit_trans(osb, handle);
1092 ocfs2_inode_unlock(inode, 1);
1095 ocfs2_rw_unlock(inode, 1);
1099 /* Release quota pointers in case we acquired them */
1100 for (qtype = 0; qtype < MAXQUOTAS; qtype++) {
1101 dqput(transfer_to[qtype]);
1102 dqput(transfer_from[qtype]);
1105 if (!status && attr->ia_valid & ATTR_MODE) {
1106 status = ocfs2_acl_chmod(inode);
1115 int ocfs2_getattr(struct vfsmount *mnt,
1116 struct dentry *dentry,
1119 struct inode *inode = dentry->d_inode;
1120 struct super_block *sb = dentry->d_inode->i_sb;
1121 struct ocfs2_super *osb = sb->s_fs_info;
1126 err = ocfs2_inode_revalidate(dentry);
1133 generic_fillattr(inode, stat);
1135 /* We set the blksize from the cluster size for performance */
1136 stat->blksize = osb->s_clustersize;
1144 int ocfs2_permission(struct inode *inode, int mask)
1150 ret = ocfs2_inode_lock(inode, NULL, 0);
1157 ret = generic_permission(inode, mask, ocfs2_check_acl);
1159 ocfs2_inode_unlock(inode, 0);
1165 static int __ocfs2_write_remove_suid(struct inode *inode,
1166 struct buffer_head *bh)
1170 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1171 struct ocfs2_dinode *di;
1173 mlog_entry("(Inode %llu, mode 0%o)\n",
1174 (unsigned long long)OCFS2_I(inode)->ip_blkno, inode->i_mode);
1176 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1177 if (IS_ERR(handle)) {
1178 ret = PTR_ERR(handle);
1183 ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), bh,
1184 OCFS2_JOURNAL_ACCESS_WRITE);
1190 inode->i_mode &= ~S_ISUID;
1191 if ((inode->i_mode & S_ISGID) && (inode->i_mode & S_IXGRP))
1192 inode->i_mode &= ~S_ISGID;
1194 di = (struct ocfs2_dinode *) bh->b_data;
1195 di->i_mode = cpu_to_le16(inode->i_mode);
1197 ret = ocfs2_journal_dirty(handle, bh);
1202 ocfs2_commit_trans(osb, handle);
1209 * Will look for holes and unwritten extents in the range starting at
1210 * pos for count bytes (inclusive).
1212 static int ocfs2_check_range_for_holes(struct inode *inode, loff_t pos,
1216 unsigned int extent_flags;
1217 u32 cpos, clusters, extent_len, phys_cpos;
1218 struct super_block *sb = inode->i_sb;
1220 cpos = pos >> OCFS2_SB(sb)->s_clustersize_bits;
1221 clusters = ocfs2_clusters_for_bytes(sb, pos + count) - cpos;
1224 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos, &extent_len,
1231 if (phys_cpos == 0 || (extent_flags & OCFS2_EXT_UNWRITTEN)) {
1236 if (extent_len > clusters)
1237 extent_len = clusters;
1239 clusters -= extent_len;
1246 static int ocfs2_write_remove_suid(struct inode *inode)
1249 struct buffer_head *bh = NULL;
1251 ret = ocfs2_read_inode_block(inode, &bh);
1257 ret = __ocfs2_write_remove_suid(inode, bh);
1264 * Allocate enough extents to cover the region starting at byte offset
1265 * start for len bytes. Existing extents are skipped, any extents
1266 * added are marked as "unwritten".
1268 static int ocfs2_allocate_unwritten_extents(struct inode *inode,
1272 u32 cpos, phys_cpos, clusters, alloc_size;
1273 u64 end = start + len;
1274 struct buffer_head *di_bh = NULL;
1276 if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1277 ret = ocfs2_read_inode_block(inode, &di_bh);
1284 * Nothing to do if the requested reservation range
1285 * fits within the inode.
1287 if (ocfs2_size_fits_inline_data(di_bh, end))
1290 ret = ocfs2_convert_inline_data_to_extents(inode, di_bh);
1298 * We consider both start and len to be inclusive.
1300 cpos = start >> OCFS2_SB(inode->i_sb)->s_clustersize_bits;
1301 clusters = ocfs2_clusters_for_bytes(inode->i_sb, start + len);
1305 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos,
1313 * Hole or existing extent len can be arbitrary, so
1314 * cap it to our own allocation request.
1316 if (alloc_size > clusters)
1317 alloc_size = clusters;
1321 * We already have an allocation at this
1322 * region so we can safely skip it.
1327 ret = __ocfs2_extend_allocation(inode, cpos, alloc_size, 1);
1336 clusters -= alloc_size;
1347 * Truncate a byte range, avoiding pages within partial clusters. This
1348 * preserves those pages for the zeroing code to write to.
1350 static void ocfs2_truncate_cluster_pages(struct inode *inode, u64 byte_start,
1353 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1355 struct address_space *mapping = inode->i_mapping;
1357 start = (loff_t)ocfs2_align_bytes_to_clusters(inode->i_sb, byte_start);
1358 end = byte_start + byte_len;
1359 end = end & ~(osb->s_clustersize - 1);
1362 unmap_mapping_range(mapping, start, end - start, 0);
1363 truncate_inode_pages_range(mapping, start, end - 1);
1367 static int ocfs2_zero_partial_clusters(struct inode *inode,
1371 u64 tmpend, end = start + len;
1372 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1373 unsigned int csize = osb->s_clustersize;
1377 * The "start" and "end" values are NOT necessarily part of
1378 * the range whose allocation is being deleted. Rather, this
1379 * is what the user passed in with the request. We must zero
1380 * partial clusters here. There's no need to worry about
1381 * physical allocation - the zeroing code knows to skip holes.
1383 mlog(0, "byte start: %llu, end: %llu\n",
1384 (unsigned long long)start, (unsigned long long)end);
1387 * If both edges are on a cluster boundary then there's no
1388 * zeroing required as the region is part of the allocation to
1391 if ((start & (csize - 1)) == 0 && (end & (csize - 1)) == 0)
1394 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1395 if (IS_ERR(handle)) {
1396 ret = PTR_ERR(handle);
1402 * We want to get the byte offset of the end of the 1st cluster.
1404 tmpend = (u64)osb->s_clustersize + (start & ~(osb->s_clustersize - 1));
1408 mlog(0, "1st range: start: %llu, tmpend: %llu\n",
1409 (unsigned long long)start, (unsigned long long)tmpend);
1411 ret = ocfs2_zero_range_for_truncate(inode, handle, start, tmpend);
1417 * This may make start and end equal, but the zeroing
1418 * code will skip any work in that case so there's no
1419 * need to catch it up here.
1421 start = end & ~(osb->s_clustersize - 1);
1423 mlog(0, "2nd range: start: %llu, end: %llu\n",
1424 (unsigned long long)start, (unsigned long long)end);
1426 ret = ocfs2_zero_range_for_truncate(inode, handle, start, end);
1431 ocfs2_commit_trans(osb, handle);
1436 static int ocfs2_remove_inode_range(struct inode *inode,
1437 struct buffer_head *di_bh, u64 byte_start,
1441 u32 trunc_start, trunc_len, cpos, phys_cpos, alloc_size;
1442 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1443 struct ocfs2_cached_dealloc_ctxt dealloc;
1444 struct address_space *mapping = inode->i_mapping;
1445 struct ocfs2_extent_tree et;
1447 ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), di_bh);
1448 ocfs2_init_dealloc_ctxt(&dealloc);
1453 if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1454 ret = ocfs2_truncate_inline(inode, di_bh, byte_start,
1455 byte_start + byte_len, 0);
1461 * There's no need to get fancy with the page cache
1462 * truncate of an inline-data inode. We're talking
1463 * about less than a page here, which will be cached
1464 * in the dinode buffer anyway.
1466 unmap_mapping_range(mapping, 0, 0, 0);
1467 truncate_inode_pages(mapping, 0);
1471 trunc_start = ocfs2_clusters_for_bytes(osb->sb, byte_start);
1472 trunc_len = (byte_start + byte_len) >> osb->s_clustersize_bits;
1473 if (trunc_len >= trunc_start)
1474 trunc_len -= trunc_start;
1478 mlog(0, "Inode: %llu, start: %llu, len: %llu, cstart: %u, clen: %u\n",
1479 (unsigned long long)OCFS2_I(inode)->ip_blkno,
1480 (unsigned long long)byte_start,
1481 (unsigned long long)byte_len, trunc_start, trunc_len);
1483 ret = ocfs2_zero_partial_clusters(inode, byte_start, byte_len);
1491 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos,
1498 if (alloc_size > trunc_len)
1499 alloc_size = trunc_len;
1501 /* Only do work for non-holes */
1502 if (phys_cpos != 0) {
1503 ret = ocfs2_remove_btree_range(inode, &et, cpos,
1504 phys_cpos, alloc_size,
1513 trunc_len -= alloc_size;
1516 ocfs2_truncate_cluster_pages(inode, byte_start, byte_len);
1519 ocfs2_schedule_truncate_log_flush(osb, 1);
1520 ocfs2_run_deallocs(osb, &dealloc);
1526 * Parts of this function taken from xfs_change_file_space()
1528 static int __ocfs2_change_file_space(struct file *file, struct inode *inode,
1529 loff_t f_pos, unsigned int cmd,
1530 struct ocfs2_space_resv *sr,
1536 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1537 struct buffer_head *di_bh = NULL;
1539 unsigned long long max_off = inode->i_sb->s_maxbytes;
1541 if (ocfs2_is_hard_readonly(osb) || ocfs2_is_soft_readonly(osb))
1544 mutex_lock(&inode->i_mutex);
1547 * This prevents concurrent writes on other nodes
1549 ret = ocfs2_rw_lock(inode, 1);
1555 ret = ocfs2_inode_lock(inode, &di_bh, 1);
1561 if (inode->i_flags & (S_IMMUTABLE|S_APPEND)) {
1563 goto out_inode_unlock;
1566 switch (sr->l_whence) {
1567 case 0: /*SEEK_SET*/
1569 case 1: /*SEEK_CUR*/
1570 sr->l_start += f_pos;
1572 case 2: /*SEEK_END*/
1573 sr->l_start += i_size_read(inode);
1577 goto out_inode_unlock;
1581 llen = sr->l_len > 0 ? sr->l_len - 1 : sr->l_len;
1584 || sr->l_start > max_off
1585 || (sr->l_start + llen) < 0
1586 || (sr->l_start + llen) > max_off) {
1588 goto out_inode_unlock;
1590 size = sr->l_start + sr->l_len;
1592 if (cmd == OCFS2_IOC_RESVSP || cmd == OCFS2_IOC_RESVSP64) {
1593 if (sr->l_len <= 0) {
1595 goto out_inode_unlock;
1599 if (file && should_remove_suid(file->f_path.dentry)) {
1600 ret = __ocfs2_write_remove_suid(inode, di_bh);
1603 goto out_inode_unlock;
1607 down_write(&OCFS2_I(inode)->ip_alloc_sem);
1609 case OCFS2_IOC_RESVSP:
1610 case OCFS2_IOC_RESVSP64:
1612 * This takes unsigned offsets, but the signed ones we
1613 * pass have been checked against overflow above.
1615 ret = ocfs2_allocate_unwritten_extents(inode, sr->l_start,
1618 case OCFS2_IOC_UNRESVSP:
1619 case OCFS2_IOC_UNRESVSP64:
1620 ret = ocfs2_remove_inode_range(inode, di_bh, sr->l_start,
1626 up_write(&OCFS2_I(inode)->ip_alloc_sem);
1629 goto out_inode_unlock;
1633 * We update c/mtime for these changes
1635 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1636 if (IS_ERR(handle)) {
1637 ret = PTR_ERR(handle);
1639 goto out_inode_unlock;
1642 if (change_size && i_size_read(inode) < size)
1643 i_size_write(inode, size);
1645 inode->i_ctime = inode->i_mtime = CURRENT_TIME;
1646 ret = ocfs2_mark_inode_dirty(handle, inode, di_bh);
1650 ocfs2_commit_trans(osb, handle);
1654 ocfs2_inode_unlock(inode, 1);
1656 ocfs2_rw_unlock(inode, 1);
1659 mutex_unlock(&inode->i_mutex);
1663 int ocfs2_change_file_space(struct file *file, unsigned int cmd,
1664 struct ocfs2_space_resv *sr)
1666 struct inode *inode = file->f_path.dentry->d_inode;
1667 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1669 if ((cmd == OCFS2_IOC_RESVSP || cmd == OCFS2_IOC_RESVSP64) &&
1670 !ocfs2_writes_unwritten_extents(osb))
1672 else if ((cmd == OCFS2_IOC_UNRESVSP || cmd == OCFS2_IOC_UNRESVSP64) &&
1673 !ocfs2_sparse_alloc(osb))
1676 if (!S_ISREG(inode->i_mode))
1679 if (!(file->f_mode & FMODE_WRITE))
1682 return __ocfs2_change_file_space(file, inode, file->f_pos, cmd, sr, 0);
1685 static long ocfs2_fallocate(struct inode *inode, int mode, loff_t offset,
1688 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1689 struct ocfs2_space_resv sr;
1690 int change_size = 1;
1692 if (!ocfs2_writes_unwritten_extents(osb))
1695 if (S_ISDIR(inode->i_mode))
1698 if (mode & FALLOC_FL_KEEP_SIZE)
1702 sr.l_start = (s64)offset;
1703 sr.l_len = (s64)len;
1705 return __ocfs2_change_file_space(NULL, inode, offset,
1706 OCFS2_IOC_RESVSP64, &sr, change_size);
1709 int ocfs2_check_range_for_refcount(struct inode *inode, loff_t pos,
1713 unsigned int extent_flags;
1714 u32 cpos, clusters, extent_len, phys_cpos;
1715 struct super_block *sb = inode->i_sb;
1717 if (!ocfs2_refcount_tree(OCFS2_SB(inode->i_sb)) ||
1718 !(OCFS2_I(inode)->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL) ||
1719 OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
1722 cpos = pos >> OCFS2_SB(sb)->s_clustersize_bits;
1723 clusters = ocfs2_clusters_for_bytes(sb, pos + count) - cpos;
1726 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos, &extent_len,
1733 if (phys_cpos && (extent_flags & OCFS2_EXT_REFCOUNTED)) {
1738 if (extent_len > clusters)
1739 extent_len = clusters;
1741 clusters -= extent_len;
1748 static int ocfs2_prepare_inode_for_refcount(struct inode *inode,
1749 loff_t pos, size_t count,
1753 struct buffer_head *di_bh = NULL;
1754 u32 cpos = pos >> OCFS2_SB(inode->i_sb)->s_clustersize_bits;
1756 ocfs2_clusters_for_bytes(inode->i_sb, pos + count) - cpos;
1758 ret = ocfs2_inode_lock(inode, &di_bh, 1);
1766 ret = ocfs2_refcount_cow(inode, di_bh, cpos, clusters, UINT_MAX);
1774 static int ocfs2_prepare_inode_for_write(struct dentry *dentry,
1781 int ret = 0, meta_level = 0;
1782 struct inode *inode = dentry->d_inode;
1783 loff_t saved_pos, end;
1786 * We start with a read level meta lock and only jump to an ex
1787 * if we need to make modifications here.
1790 ret = ocfs2_inode_lock(inode, NULL, meta_level);
1797 /* Clear suid / sgid if necessary. We do this here
1798 * instead of later in the write path because
1799 * remove_suid() calls ->setattr without any hint that
1800 * we may have already done our cluster locking. Since
1801 * ocfs2_setattr() *must* take cluster locks to
1802 * proceeed, this will lead us to recursively lock the
1803 * inode. There's also the dinode i_size state which
1804 * can be lost via setattr during extending writes (we
1805 * set inode->i_size at the end of a write. */
1806 if (should_remove_suid(dentry)) {
1807 if (meta_level == 0) {
1808 ocfs2_inode_unlock(inode, meta_level);
1813 ret = ocfs2_write_remove_suid(inode);
1820 /* work on a copy of ppos until we're sure that we won't have
1821 * to recalculate it due to relocking. */
1823 saved_pos = i_size_read(inode);
1824 mlog(0, "O_APPEND: inode->i_size=%llu\n", saved_pos);
1829 end = saved_pos + count;
1831 ret = ocfs2_check_range_for_refcount(inode, saved_pos, count);
1833 ocfs2_inode_unlock(inode, meta_level);
1836 ret = ocfs2_prepare_inode_for_refcount(inode,
1852 * Skip the O_DIRECT checks if we don't need
1855 if (!direct_io || !(*direct_io))
1859 * There's no sane way to do direct writes to an inode
1862 if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1868 * Allowing concurrent direct writes means
1869 * i_size changes wouldn't be synchronized, so
1870 * one node could wind up truncating another
1873 if (end > i_size_read(inode)) {
1879 * We don't fill holes during direct io, so
1880 * check for them here. If any are found, the
1881 * caller will have to retake some cluster
1882 * locks and initiate the io as buffered.
1884 ret = ocfs2_check_range_for_holes(inode, saved_pos, count);
1897 if (meta_level >= 0)
1898 ocfs2_inode_unlock(inode, meta_level);
1904 static ssize_t ocfs2_file_aio_write(struct kiocb *iocb,
1905 const struct iovec *iov,
1906 unsigned long nr_segs,
1909 int ret, direct_io, appending, rw_level, have_alloc_sem = 0;
1910 int can_do_direct, has_refcount = 0;
1911 ssize_t written = 0;
1912 size_t ocount; /* original count */
1913 size_t count; /* after file limit checks */
1914 loff_t old_size, *ppos = &iocb->ki_pos;
1916 struct file *file = iocb->ki_filp;
1917 struct inode *inode = file->f_path.dentry->d_inode;
1918 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1920 mlog_entry("(0x%p, %u, '%.*s')\n", file,
1921 (unsigned int)nr_segs,
1922 file->f_path.dentry->d_name.len,
1923 file->f_path.dentry->d_name.name);
1925 if (iocb->ki_left == 0)
1928 vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
1930 appending = file->f_flags & O_APPEND ? 1 : 0;
1931 direct_io = file->f_flags & O_DIRECT ? 1 : 0;
1933 mutex_lock(&inode->i_mutex);
1936 /* to match setattr's i_mutex -> i_alloc_sem -> rw_lock ordering */
1938 down_read(&inode->i_alloc_sem);
1942 /* concurrent O_DIRECT writes are allowed */
1943 rw_level = !direct_io;
1944 ret = ocfs2_rw_lock(inode, rw_level);
1950 can_do_direct = direct_io;
1951 ret = ocfs2_prepare_inode_for_write(file->f_path.dentry, ppos,
1952 iocb->ki_left, appending,
1953 &can_do_direct, &has_refcount);
1960 * We can't complete the direct I/O as requested, fall back to
1963 if (direct_io && !can_do_direct) {
1964 ocfs2_rw_unlock(inode, rw_level);
1965 up_read(&inode->i_alloc_sem);
1975 * To later detect whether a journal commit for sync writes is
1976 * necessary, we sample i_size, and cluster count here.
1978 old_size = i_size_read(inode);
1979 old_clusters = OCFS2_I(inode)->ip_clusters;
1981 /* communicate with ocfs2_dio_end_io */
1982 ocfs2_iocb_set_rw_locked(iocb, rw_level);
1985 ret = generic_segment_checks(iov, &nr_segs, &ocount,
1991 ret = generic_write_checks(file, ppos, &count,
1992 S_ISBLK(inode->i_mode));
1996 written = generic_file_direct_write(iocb, iov, &nr_segs, *ppos,
1997 ppos, count, ocount);
2000 * direct write may have instantiated a few
2001 * blocks outside i_size. Trim these off again.
2002 * Don't need i_size_read because we hold i_mutex.
2004 if (*ppos + count > inode->i_size)
2005 vmtruncate(inode, inode->i_size);
2010 written = __generic_file_aio_write(iocb, iov, nr_segs, ppos);
2014 /* buffered aio wouldn't have proper lock coverage today */
2015 BUG_ON(ret == -EIOCBQUEUED && !(file->f_flags & O_DIRECT));
2017 if (((file->f_flags & O_DSYNC) && !direct_io) || IS_SYNC(inode) ||
2018 ((file->f_flags & O_DIRECT) && has_refcount)) {
2019 ret = filemap_fdatawrite_range(file->f_mapping, pos,
2024 if (!ret && (old_size != i_size_read(inode) ||
2025 old_clusters != OCFS2_I(inode)->ip_clusters ||
2027 ret = jbd2_journal_force_commit(osb->journal->j_journal);
2033 ret = filemap_fdatawait_range(file->f_mapping, pos,
2038 * deep in g_f_a_w_n()->ocfs2_direct_IO we pass in a ocfs2_dio_end_io
2039 * function pointer which is called when o_direct io completes so that
2040 * it can unlock our rw lock. (it's the clustered equivalent of
2041 * i_alloc_sem; protects truncate from racing with pending ios).
2042 * Unfortunately there are error cases which call end_io and others
2043 * that don't. so we don't have to unlock the rw_lock if either an
2044 * async dio is going to do it in the future or an end_io after an
2045 * error has already done it.
2047 if ((ret == -EIOCBQUEUED) || (!ocfs2_iocb_is_rw_locked(iocb))) {
2054 ocfs2_rw_unlock(inode, rw_level);
2058 up_read(&inode->i_alloc_sem);
2060 mutex_unlock(&inode->i_mutex);
2068 static int ocfs2_splice_to_file(struct pipe_inode_info *pipe,
2070 struct splice_desc *sd)
2074 ret = ocfs2_prepare_inode_for_write(out->f_path.dentry, &sd->pos,
2075 sd->total_len, 0, NULL, NULL);
2081 return splice_from_pipe_feed(pipe, sd, pipe_to_file);
2084 static ssize_t ocfs2_file_splice_write(struct pipe_inode_info *pipe,
2091 struct address_space *mapping = out->f_mapping;
2092 struct inode *inode = mapping->host;
2093 struct splice_desc sd = {
2100 mlog_entry("(0x%p, 0x%p, %u, '%.*s')\n", out, pipe,
2102 out->f_path.dentry->d_name.len,
2103 out->f_path.dentry->d_name.name);
2106 mutex_lock_nested(&pipe->inode->i_mutex, I_MUTEX_PARENT);
2108 splice_from_pipe_begin(&sd);
2110 ret = splice_from_pipe_next(pipe, &sd);
2114 mutex_lock_nested(&inode->i_mutex, I_MUTEX_CHILD);
2115 ret = ocfs2_rw_lock(inode, 1);
2119 ret = ocfs2_splice_to_file(pipe, out, &sd);
2120 ocfs2_rw_unlock(inode, 1);
2122 mutex_unlock(&inode->i_mutex);
2124 splice_from_pipe_end(pipe, &sd);
2127 mutex_unlock(&pipe->inode->i_mutex);
2130 ret = sd.num_spliced;
2133 unsigned long nr_pages;
2136 nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
2138 err = generic_write_sync(out, *ppos, ret);
2144 balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
2151 static ssize_t ocfs2_file_splice_read(struct file *in,
2153 struct pipe_inode_info *pipe,
2157 int ret = 0, lock_level = 0;
2158 struct inode *inode = in->f_path.dentry->d_inode;
2160 mlog_entry("(0x%p, 0x%p, %u, '%.*s')\n", in, pipe,
2162 in->f_path.dentry->d_name.len,
2163 in->f_path.dentry->d_name.name);
2166 * See the comment in ocfs2_file_aio_read()
2168 ret = ocfs2_inode_lock_atime(inode, in->f_vfsmnt, &lock_level);
2173 ocfs2_inode_unlock(inode, lock_level);
2175 ret = generic_file_splice_read(in, ppos, pipe, len, flags);
2182 static ssize_t ocfs2_file_aio_read(struct kiocb *iocb,
2183 const struct iovec *iov,
2184 unsigned long nr_segs,
2187 int ret = 0, rw_level = -1, have_alloc_sem = 0, lock_level = 0;
2188 struct file *filp = iocb->ki_filp;
2189 struct inode *inode = filp->f_path.dentry->d_inode;
2191 mlog_entry("(0x%p, %u, '%.*s')\n", filp,
2192 (unsigned int)nr_segs,
2193 filp->f_path.dentry->d_name.len,
2194 filp->f_path.dentry->d_name.name);
2203 * buffered reads protect themselves in ->readpage(). O_DIRECT reads
2204 * need locks to protect pending reads from racing with truncate.
2206 if (filp->f_flags & O_DIRECT) {
2207 down_read(&inode->i_alloc_sem);
2210 ret = ocfs2_rw_lock(inode, 0);
2216 /* communicate with ocfs2_dio_end_io */
2217 ocfs2_iocb_set_rw_locked(iocb, rw_level);
2221 * We're fine letting folks race truncates and extending
2222 * writes with read across the cluster, just like they can
2223 * locally. Hence no rw_lock during read.
2225 * Take and drop the meta data lock to update inode fields
2226 * like i_size. This allows the checks down below
2227 * generic_file_aio_read() a chance of actually working.
2229 ret = ocfs2_inode_lock_atime(inode, filp->f_vfsmnt, &lock_level);
2234 ocfs2_inode_unlock(inode, lock_level);
2236 ret = generic_file_aio_read(iocb, iov, nr_segs, iocb->ki_pos);
2238 mlog(0, "generic_file_aio_read returned -EINVAL\n");
2240 /* buffered aio wouldn't have proper lock coverage today */
2241 BUG_ON(ret == -EIOCBQUEUED && !(filp->f_flags & O_DIRECT));
2243 /* see ocfs2_file_aio_write */
2244 if (ret == -EIOCBQUEUED || !ocfs2_iocb_is_rw_locked(iocb)) {
2251 up_read(&inode->i_alloc_sem);
2253 ocfs2_rw_unlock(inode, rw_level);
2259 const struct inode_operations ocfs2_file_iops = {
2260 .setattr = ocfs2_setattr,
2261 .getattr = ocfs2_getattr,
2262 .permission = ocfs2_permission,
2263 .setxattr = generic_setxattr,
2264 .getxattr = generic_getxattr,
2265 .listxattr = ocfs2_listxattr,
2266 .removexattr = generic_removexattr,
2267 .fallocate = ocfs2_fallocate,
2268 .fiemap = ocfs2_fiemap,
2271 const struct inode_operations ocfs2_special_file_iops = {
2272 .setattr = ocfs2_setattr,
2273 .getattr = ocfs2_getattr,
2274 .permission = ocfs2_permission,
2278 * Other than ->lock, keep ocfs2_fops and ocfs2_dops in sync with
2279 * ocfs2_fops_no_plocks and ocfs2_dops_no_plocks!
2281 const struct file_operations ocfs2_fops = {
2282 .llseek = generic_file_llseek,
2283 .read = do_sync_read,
2284 .write = do_sync_write,
2286 .fsync = ocfs2_sync_file,
2287 .release = ocfs2_file_release,
2288 .open = ocfs2_file_open,
2289 .aio_read = ocfs2_file_aio_read,
2290 .aio_write = ocfs2_file_aio_write,
2291 .unlocked_ioctl = ocfs2_ioctl,
2292 #ifdef CONFIG_COMPAT
2293 .compat_ioctl = ocfs2_compat_ioctl,
2296 .flock = ocfs2_flock,
2297 .splice_read = ocfs2_file_splice_read,
2298 .splice_write = ocfs2_file_splice_write,
2301 const struct file_operations ocfs2_dops = {
2302 .llseek = generic_file_llseek,
2303 .read = generic_read_dir,
2304 .readdir = ocfs2_readdir,
2305 .fsync = ocfs2_sync_file,
2306 .release = ocfs2_dir_release,
2307 .open = ocfs2_dir_open,
2308 .unlocked_ioctl = ocfs2_ioctl,
2309 #ifdef CONFIG_COMPAT
2310 .compat_ioctl = ocfs2_compat_ioctl,
2313 .flock = ocfs2_flock,
2317 * POSIX-lockless variants of our file_operations.
2319 * These will be used if the underlying cluster stack does not support
2320 * posix file locking, if the user passes the "localflocks" mount
2321 * option, or if we have a local-only fs.
2323 * ocfs2_flock is in here because all stacks handle UNIX file locks,
2324 * so we still want it in the case of no stack support for
2325 * plocks. Internally, it will do the right thing when asked to ignore
2328 const struct file_operations ocfs2_fops_no_plocks = {
2329 .llseek = generic_file_llseek,
2330 .read = do_sync_read,
2331 .write = do_sync_write,
2333 .fsync = ocfs2_sync_file,
2334 .release = ocfs2_file_release,
2335 .open = ocfs2_file_open,
2336 .aio_read = ocfs2_file_aio_read,
2337 .aio_write = ocfs2_file_aio_write,
2338 .unlocked_ioctl = ocfs2_ioctl,
2339 #ifdef CONFIG_COMPAT
2340 .compat_ioctl = ocfs2_compat_ioctl,
2342 .flock = ocfs2_flock,
2343 .splice_read = ocfs2_file_splice_read,
2344 .splice_write = ocfs2_file_splice_write,
2347 const struct file_operations ocfs2_dops_no_plocks = {
2348 .llseek = generic_file_llseek,
2349 .read = generic_read_dir,
2350 .readdir = ocfs2_readdir,
2351 .fsync = ocfs2_sync_file,
2352 .release = ocfs2_dir_release,
2353 .open = ocfs2_dir_open,
2354 .unlocked_ioctl = ocfs2_ioctl,
2355 #ifdef CONFIG_COMPAT
2356 .compat_ioctl = ocfs2_compat_ioctl,
2358 .flock = ocfs2_flock,