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/pipe_fs_i.h>
35 #include <linux/mount.h>
36 #include <linux/writeback.h>
38 #define MLOG_MASK_PREFIX ML_INODE
39 #include <cluster/masklog.h>
47 #include "extent_map.h"
57 #include "buffer_head_io.h"
59 static int ocfs2_sync_inode(struct inode *inode)
61 filemap_fdatawrite(inode->i_mapping);
62 return sync_mapping_buffers(inode->i_mapping);
65 static int ocfs2_file_open(struct inode *inode, struct file *file)
68 int mode = file->f_flags;
69 struct ocfs2_inode_info *oi = OCFS2_I(inode);
71 mlog_entry("(0x%p, 0x%p, '%.*s')\n", inode, file,
72 file->f_path.dentry->d_name.len, file->f_path.dentry->d_name.name);
74 spin_lock(&oi->ip_lock);
76 /* Check that the inode hasn't been wiped from disk by another
77 * node. If it hasn't then we're safe as long as we hold the
78 * spin lock until our increment of open count. */
79 if (OCFS2_I(inode)->ip_flags & OCFS2_INODE_DELETED) {
80 spin_unlock(&oi->ip_lock);
87 oi->ip_flags |= OCFS2_INODE_OPEN_DIRECT;
90 spin_unlock(&oi->ip_lock);
97 static int ocfs2_file_release(struct inode *inode, struct file *file)
99 struct ocfs2_inode_info *oi = OCFS2_I(inode);
101 mlog_entry("(0x%p, 0x%p, '%.*s')\n", inode, file,
102 file->f_path.dentry->d_name.len,
103 file->f_path.dentry->d_name.name);
105 spin_lock(&oi->ip_lock);
106 if (!--oi->ip_open_count)
107 oi->ip_flags &= ~OCFS2_INODE_OPEN_DIRECT;
108 spin_unlock(&oi->ip_lock);
115 static int ocfs2_sync_file(struct file *file,
116 struct dentry *dentry,
121 struct inode *inode = dentry->d_inode;
122 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
124 mlog_entry("(0x%p, 0x%p, %d, '%.*s')\n", file, dentry, datasync,
125 dentry->d_name.len, dentry->d_name.name);
127 err = ocfs2_sync_inode(dentry->d_inode);
131 journal = osb->journal->j_journal;
132 err = journal_force_commit(journal);
137 return (err < 0) ? -EIO : 0;
140 int ocfs2_should_update_atime(struct inode *inode,
141 struct vfsmount *vfsmnt)
144 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
146 if (ocfs2_is_hard_readonly(osb) || ocfs2_is_soft_readonly(osb))
149 if ((inode->i_flags & S_NOATIME) ||
150 ((inode->i_sb->s_flags & MS_NODIRATIME) && S_ISDIR(inode->i_mode)))
154 * We can be called with no vfsmnt structure - NFSD will
157 * Note that our action here is different than touch_atime() -
158 * if we can't tell whether this is a noatime mount, then we
159 * don't know whether to trust the value of s_atime_quantum.
164 if ((vfsmnt->mnt_flags & MNT_NOATIME) ||
165 ((vfsmnt->mnt_flags & MNT_NODIRATIME) && S_ISDIR(inode->i_mode)))
168 if (vfsmnt->mnt_flags & MNT_RELATIME) {
169 if ((timespec_compare(&inode->i_atime, &inode->i_mtime) <= 0) ||
170 (timespec_compare(&inode->i_atime, &inode->i_ctime) <= 0))
177 if ((now.tv_sec - inode->i_atime.tv_sec <= osb->s_atime_quantum))
183 int ocfs2_update_inode_atime(struct inode *inode,
184 struct buffer_head *bh)
187 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
192 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
193 if (handle == NULL) {
199 inode->i_atime = CURRENT_TIME;
200 ret = ocfs2_mark_inode_dirty(handle, inode, bh);
204 ocfs2_commit_trans(OCFS2_SB(inode->i_sb), handle);
210 int ocfs2_set_inode_size(handle_t *handle,
212 struct buffer_head *fe_bh,
218 i_size_write(inode, new_i_size);
219 inode->i_blocks = ocfs2_align_bytes_to_sectors(new_i_size);
220 inode->i_ctime = inode->i_mtime = CURRENT_TIME;
222 status = ocfs2_mark_inode_dirty(handle, inode, fe_bh);
233 static int ocfs2_simple_size_update(struct inode *inode,
234 struct buffer_head *di_bh,
238 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
239 handle_t *handle = NULL;
241 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
242 if (handle == NULL) {
248 ret = ocfs2_set_inode_size(handle, inode, di_bh,
253 ocfs2_commit_trans(osb, handle);
258 static int ocfs2_orphan_for_truncate(struct ocfs2_super *osb,
260 struct buffer_head *fe_bh,
268 /* TODO: This needs to actually orphan the inode in this
271 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
272 if (IS_ERR(handle)) {
273 status = PTR_ERR(handle);
278 status = ocfs2_set_inode_size(handle, inode, fe_bh, new_i_size);
282 ocfs2_commit_trans(osb, handle);
288 static int ocfs2_truncate_file(struct inode *inode,
289 struct buffer_head *di_bh,
293 struct ocfs2_dinode *fe = NULL;
294 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
295 struct ocfs2_truncate_context *tc = NULL;
297 mlog_entry("(inode = %llu, new_i_size = %llu\n",
298 (unsigned long long)OCFS2_I(inode)->ip_blkno,
299 (unsigned long long)new_i_size);
301 truncate_inode_pages(inode->i_mapping, new_i_size);
303 fe = (struct ocfs2_dinode *) di_bh->b_data;
304 if (!OCFS2_IS_VALID_DINODE(fe)) {
305 OCFS2_RO_ON_INVALID_DINODE(inode->i_sb, fe);
310 mlog_bug_on_msg(le64_to_cpu(fe->i_size) != i_size_read(inode),
311 "Inode %llu, inode i_size = %lld != di "
312 "i_size = %llu, i_flags = 0x%x\n",
313 (unsigned long long)OCFS2_I(inode)->ip_blkno,
315 (unsigned long long)le64_to_cpu(fe->i_size),
316 le32_to_cpu(fe->i_flags));
318 if (new_i_size > le64_to_cpu(fe->i_size)) {
319 mlog(0, "asked to truncate file with size (%llu) to size (%llu)!\n",
320 (unsigned long long)le64_to_cpu(fe->i_size),
321 (unsigned long long)new_i_size);
327 mlog(0, "inode %llu, i_size = %llu, new_i_size = %llu\n",
328 (unsigned long long)le64_to_cpu(fe->i_blkno),
329 (unsigned long long)le64_to_cpu(fe->i_size),
330 (unsigned long long)new_i_size);
332 /* lets handle the simple truncate cases before doing any more
333 * cluster locking. */
334 if (new_i_size == le64_to_cpu(fe->i_size))
337 /* This forces other nodes to sync and drop their pages. Do
338 * this even if we have a truncate without allocation change -
339 * ocfs2 cluster sizes can be much greater than page size, so
340 * we have to truncate them anyway. */
341 status = ocfs2_data_lock(inode, 1);
346 ocfs2_data_unlock(inode, 1);
348 /* alright, we're going to need to do a full blown alloc size
349 * change. Orphan the inode so that recovery can complete the
350 * truncate if necessary. This does the task of marking
352 status = ocfs2_orphan_for_truncate(osb, inode, di_bh, new_i_size);
358 status = ocfs2_prepare_truncate(osb, inode, di_bh, &tc);
364 status = ocfs2_commit_truncate(osb, inode, di_bh, tc);
370 /* TODO: orphan dir cleanup here. */
378 * extend allocation only here.
379 * we'll update all the disk stuff, and oip->alloc_size
381 * expect stuff to be locked, a transaction started and enough data /
382 * metadata reservations in the contexts.
384 * Will return -EAGAIN, and a reason if a restart is needed.
385 * If passed in, *reason will always be set, even in error.
387 int ocfs2_do_extend_allocation(struct ocfs2_super *osb,
391 struct buffer_head *fe_bh,
393 struct ocfs2_alloc_context *data_ac,
394 struct ocfs2_alloc_context *meta_ac,
395 enum ocfs2_alloc_restarted *reason_ret)
399 struct ocfs2_dinode *fe = (struct ocfs2_dinode *) fe_bh->b_data;
400 enum ocfs2_alloc_restarted reason = RESTART_NONE;
401 u32 bit_off, num_bits;
404 BUG_ON(!clusters_to_add);
406 free_extents = ocfs2_num_free_extents(osb, inode, fe);
407 if (free_extents < 0) {
408 status = free_extents;
413 /* there are two cases which could cause us to EAGAIN in the
414 * we-need-more-metadata case:
415 * 1) we haven't reserved *any*
416 * 2) we are so fragmented, we've needed to add metadata too
418 if (!free_extents && !meta_ac) {
419 mlog(0, "we haven't reserved any metadata!\n");
421 reason = RESTART_META;
423 } else if ((!free_extents)
424 && (ocfs2_alloc_context_bits_left(meta_ac)
425 < ocfs2_extend_meta_needed(fe))) {
426 mlog(0, "filesystem is really fragmented...\n");
428 reason = RESTART_META;
432 status = ocfs2_claim_clusters(osb, handle, data_ac, 1,
433 &bit_off, &num_bits);
435 if (status != -ENOSPC)
440 BUG_ON(num_bits > clusters_to_add);
442 /* reserve our write early -- insert_extent may update the inode */
443 status = ocfs2_journal_access(handle, inode, fe_bh,
444 OCFS2_JOURNAL_ACCESS_WRITE);
450 block = ocfs2_clusters_to_blocks(osb->sb, bit_off);
451 mlog(0, "Allocating %u clusters at block %u for inode %llu\n",
452 num_bits, bit_off, (unsigned long long)OCFS2_I(inode)->ip_blkno);
453 status = ocfs2_insert_extent(osb, handle, inode, fe_bh,
454 *logical_offset, block, num_bits,
461 status = ocfs2_journal_dirty(handle, fe_bh);
467 clusters_to_add -= num_bits;
468 *logical_offset += num_bits;
470 if (clusters_to_add) {
471 mlog(0, "need to alloc once more, clusters = %u, wanted = "
472 "%u\n", fe->i_clusters, clusters_to_add);
474 reason = RESTART_TRANS;
480 *reason_ret = reason;
485 * For a given allocation, determine which allocators will need to be
486 * accessed, and lock them, reserving the appropriate number of bits.
488 * Called from ocfs2_extend_allocation() for file systems which don't
489 * support holes, and from ocfs2_write() for file systems which
490 * understand sparse inodes.
492 int ocfs2_lock_allocators(struct inode *inode, struct ocfs2_dinode *di,
494 struct ocfs2_alloc_context **data_ac,
495 struct ocfs2_alloc_context **meta_ac)
497 int ret, num_free_extents;
498 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
503 mlog(0, "extend inode %llu, i_size = %lld, di->i_clusters = %u, "
504 "clusters_to_add = %u\n",
505 (unsigned long long)OCFS2_I(inode)->ip_blkno, i_size_read(inode),
506 le32_to_cpu(di->i_clusters), clusters_to_add);
508 num_free_extents = ocfs2_num_free_extents(osb, inode, di);
509 if (num_free_extents < 0) {
510 ret = num_free_extents;
516 * Sparse allocation file systems need to be more conservative
517 * with reserving room for expansion - the actual allocation
518 * happens while we've got a journal handle open so re-taking
519 * a cluster lock (because we ran out of room for another
520 * extent) will violate ordering rules.
522 * Most of the time we'll only be seeing this 1 cluster at a time
525 if (!num_free_extents ||
526 (ocfs2_sparse_alloc(osb) && num_free_extents < clusters_to_add)) {
527 ret = ocfs2_reserve_new_metadata(osb, di, meta_ac);
535 ret = ocfs2_reserve_clusters(osb, clusters_to_add, data_ac);
545 ocfs2_free_alloc_context(*meta_ac);
550 * We cannot have an error and a non null *data_ac.
557 static int ocfs2_extend_allocation(struct inode *inode,
561 int restart_func = 0;
562 int drop_alloc_sem = 0;
564 u32 prev_clusters, logical_start;
565 struct buffer_head *bh = NULL;
566 struct ocfs2_dinode *fe = NULL;
567 handle_t *handle = NULL;
568 struct ocfs2_alloc_context *data_ac = NULL;
569 struct ocfs2_alloc_context *meta_ac = NULL;
570 enum ocfs2_alloc_restarted why;
571 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
573 mlog_entry("(clusters_to_add = %u)\n", clusters_to_add);
576 * This function only exists for file systems which don't
579 BUG_ON(ocfs2_sparse_alloc(osb));
581 status = ocfs2_read_block(osb, OCFS2_I(inode)->ip_blkno, &bh,
582 OCFS2_BH_CACHED, inode);
588 fe = (struct ocfs2_dinode *) bh->b_data;
589 if (!OCFS2_IS_VALID_DINODE(fe)) {
590 OCFS2_RO_ON_INVALID_DINODE(inode->i_sb, fe);
595 logical_start = OCFS2_I(inode)->ip_clusters;
598 BUG_ON(le32_to_cpu(fe->i_clusters) != OCFS2_I(inode)->ip_clusters);
600 /* blocks peope in read/write from reading our allocation
601 * until we're done changing it. We depend on i_mutex to block
602 * other extend/truncate calls while we're here. Ordering wrt
603 * start_trans is important here -- always do it before! */
604 down_write(&OCFS2_I(inode)->ip_alloc_sem);
607 status = ocfs2_lock_allocators(inode, fe, clusters_to_add, &data_ac,
614 credits = ocfs2_calc_extend_credits(osb->sb, fe, clusters_to_add);
615 handle = ocfs2_start_trans(osb, credits);
616 if (IS_ERR(handle)) {
617 status = PTR_ERR(handle);
623 restarted_transaction:
624 /* reserve a write to the file entry early on - that we if we
625 * run out of credits in the allocation path, we can still
627 status = ocfs2_journal_access(handle, inode, bh,
628 OCFS2_JOURNAL_ACCESS_WRITE);
634 prev_clusters = OCFS2_I(inode)->ip_clusters;
636 status = ocfs2_do_extend_allocation(osb,
645 if ((status < 0) && (status != -EAGAIN)) {
646 if (status != -ENOSPC)
651 status = ocfs2_journal_dirty(handle, bh);
657 spin_lock(&OCFS2_I(inode)->ip_lock);
658 clusters_to_add -= (OCFS2_I(inode)->ip_clusters - prev_clusters);
659 spin_unlock(&OCFS2_I(inode)->ip_lock);
661 if (why != RESTART_NONE && clusters_to_add) {
662 if (why == RESTART_META) {
663 mlog(0, "restarting function.\n");
666 BUG_ON(why != RESTART_TRANS);
668 mlog(0, "restarting transaction.\n");
669 /* TODO: This can be more intelligent. */
670 credits = ocfs2_calc_extend_credits(osb->sb,
673 status = ocfs2_extend_trans(handle, credits);
675 /* handle still has to be committed at
681 goto restarted_transaction;
685 mlog(0, "fe: i_clusters = %u, i_size=%llu\n",
686 fe->i_clusters, (unsigned long long)fe->i_size);
687 mlog(0, "inode: ip_clusters=%u, i_size=%lld\n",
688 OCFS2_I(inode)->ip_clusters, i_size_read(inode));
691 if (drop_alloc_sem) {
692 up_write(&OCFS2_I(inode)->ip_alloc_sem);
696 ocfs2_commit_trans(osb, handle);
700 ocfs2_free_alloc_context(data_ac);
704 ocfs2_free_alloc_context(meta_ac);
707 if ((!status) && restart_func) {
720 /* Some parts of this taken from generic_cont_expand, which turned out
721 * to be too fragile to do exactly what we need without us having to
722 * worry about recursive locking in ->prepare_write() and
723 * ->commit_write(). */
724 static int ocfs2_write_zero_page(struct inode *inode,
727 struct address_space *mapping = inode->i_mapping;
731 handle_t *handle = NULL;
734 offset = (size & (PAGE_CACHE_SIZE-1)); /* Within page */
735 /* ugh. in prepare/commit_write, if from==to==start of block, we
736 ** skip the prepare. make sure we never send an offset for the start
739 if ((offset & (inode->i_sb->s_blocksize - 1)) == 0) {
742 index = size >> PAGE_CACHE_SHIFT;
744 page = grab_cache_page(mapping, index);
751 ret = ocfs2_prepare_write_nolock(inode, page, offset, offset);
757 if (ocfs2_should_order_data(inode)) {
758 handle = ocfs2_start_walk_page_trans(inode, page, offset,
760 if (IS_ERR(handle)) {
761 ret = PTR_ERR(handle);
767 /* must not update i_size! */
768 ret = block_commit_write(page, offset, offset);
775 ocfs2_commit_trans(OCFS2_SB(inode->i_sb), handle);
778 page_cache_release(page);
783 static int ocfs2_zero_extend(struct inode *inode,
788 struct super_block *sb = inode->i_sb;
790 start_off = ocfs2_align_bytes_to_blocks(sb, i_size_read(inode));
791 while (start_off < zero_to_size) {
792 ret = ocfs2_write_zero_page(inode, start_off);
798 start_off += sb->s_blocksize;
801 * Very large extends have the potential to lock up
802 * the cpu for extended periods of time.
812 * A tail_to_skip value > 0 indicates that we're being called from
813 * ocfs2_file_aio_write(). This has the following implications:
815 * - we don't want to update i_size
816 * - di_bh will be NULL, which is fine because it's only used in the
817 * case where we want to update i_size.
818 * - ocfs2_zero_extend() will then only be filling the hole created
819 * between i_size and the start of the write.
821 static int ocfs2_extend_file(struct inode *inode,
822 struct buffer_head *di_bh,
827 u32 clusters_to_add = 0;
829 BUG_ON(!tail_to_skip && !di_bh);
831 /* setattr sometimes calls us like this. */
835 if (i_size_read(inode) == new_i_size)
837 BUG_ON(new_i_size < i_size_read(inode));
839 if (ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb))) {
840 BUG_ON(tail_to_skip != 0);
841 goto out_update_size;
844 clusters_to_add = ocfs2_clusters_for_bytes(inode->i_sb, new_i_size) -
845 OCFS2_I(inode)->ip_clusters;
848 * protect the pages that ocfs2_zero_extend is going to be
849 * pulling into the page cache.. we do this before the
850 * metadata extend so that we don't get into the situation
851 * where we've extended the metadata but can't get the data
854 ret = ocfs2_data_lock(inode, 1);
860 if (clusters_to_add) {
861 ret = ocfs2_extend_allocation(inode, clusters_to_add);
869 * Call this even if we don't add any clusters to the tree. We
870 * still need to zero the area between the old i_size and the
873 ret = ocfs2_zero_extend(inode, (u64)new_i_size - tail_to_skip);
881 /* We're being called from ocfs2_setattr() which wants
882 * us to update i_size */
883 ret = ocfs2_simple_size_update(inode, di_bh, new_i_size);
889 if (!ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)))
890 ocfs2_data_unlock(inode, 1);
896 int ocfs2_setattr(struct dentry *dentry, struct iattr *attr)
898 int status = 0, size_change;
899 struct inode *inode = dentry->d_inode;
900 struct super_block *sb = inode->i_sb;
901 struct ocfs2_super *osb = OCFS2_SB(sb);
902 struct buffer_head *bh = NULL;
903 handle_t *handle = NULL;
905 mlog_entry("(0x%p, '%.*s')\n", dentry,
906 dentry->d_name.len, dentry->d_name.name);
908 if (attr->ia_valid & ATTR_MODE)
909 mlog(0, "mode change: %d\n", attr->ia_mode);
910 if (attr->ia_valid & ATTR_UID)
911 mlog(0, "uid change: %d\n", attr->ia_uid);
912 if (attr->ia_valid & ATTR_GID)
913 mlog(0, "gid change: %d\n", attr->ia_gid);
914 if (attr->ia_valid & ATTR_SIZE)
915 mlog(0, "size change...\n");
916 if (attr->ia_valid & (ATTR_ATIME | ATTR_MTIME | ATTR_CTIME))
917 mlog(0, "time change...\n");
919 #define OCFS2_VALID_ATTRS (ATTR_ATIME | ATTR_MTIME | ATTR_CTIME | ATTR_SIZE \
920 | ATTR_GID | ATTR_UID | ATTR_MODE)
921 if (!(attr->ia_valid & OCFS2_VALID_ATTRS)) {
922 mlog(0, "can't handle attrs: 0x%x\n", attr->ia_valid);
926 status = inode_change_ok(inode, attr);
930 size_change = S_ISREG(inode->i_mode) && attr->ia_valid & ATTR_SIZE;
932 status = ocfs2_rw_lock(inode, 1);
939 status = ocfs2_meta_lock(inode, &bh, 1);
941 if (status != -ENOENT)
946 if (size_change && attr->ia_size != i_size_read(inode)) {
947 if (i_size_read(inode) > attr->ia_size)
948 status = ocfs2_truncate_file(inode, bh, attr->ia_size);
950 status = ocfs2_extend_file(inode, bh, attr->ia_size, 0);
952 if (status != -ENOSPC)
959 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
960 if (IS_ERR(handle)) {
961 status = PTR_ERR(handle);
966 status = inode_setattr(inode, attr);
972 status = ocfs2_mark_inode_dirty(handle, inode, bh);
977 ocfs2_commit_trans(osb, handle);
979 ocfs2_meta_unlock(inode, 1);
982 ocfs2_rw_unlock(inode, 1);
991 int ocfs2_getattr(struct vfsmount *mnt,
992 struct dentry *dentry,
995 struct inode *inode = dentry->d_inode;
996 struct super_block *sb = dentry->d_inode->i_sb;
997 struct ocfs2_super *osb = sb->s_fs_info;
1002 err = ocfs2_inode_revalidate(dentry);
1009 generic_fillattr(inode, stat);
1011 /* We set the blksize from the cluster size for performance */
1012 stat->blksize = osb->s_clustersize;
1020 int ocfs2_permission(struct inode *inode, int mask, struct nameidata *nd)
1026 ret = ocfs2_meta_lock(inode, NULL, 0);
1033 ret = generic_permission(inode, mask, NULL);
1035 ocfs2_meta_unlock(inode, 0);
1041 static int ocfs2_write_remove_suid(struct inode *inode)
1044 struct buffer_head *bh = NULL;
1045 struct ocfs2_inode_info *oi = OCFS2_I(inode);
1047 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1048 struct ocfs2_dinode *di;
1050 mlog_entry("(Inode %llu, mode 0%o)\n",
1051 (unsigned long long)oi->ip_blkno, inode->i_mode);
1053 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1054 if (handle == NULL) {
1060 ret = ocfs2_read_block(osb, oi->ip_blkno, &bh, OCFS2_BH_CACHED, inode);
1066 ret = ocfs2_journal_access(handle, inode, bh,
1067 OCFS2_JOURNAL_ACCESS_WRITE);
1073 inode->i_mode &= ~S_ISUID;
1074 if ((inode->i_mode & S_ISGID) && (inode->i_mode & S_IXGRP))
1075 inode->i_mode &= ~S_ISGID;
1077 di = (struct ocfs2_dinode *) bh->b_data;
1078 di->i_mode = cpu_to_le16(inode->i_mode);
1080 ret = ocfs2_journal_dirty(handle, bh);
1086 ocfs2_commit_trans(osb, handle);
1093 * Will look for holes and unwritten extents in the range starting at
1094 * pos for count bytes (inclusive).
1096 static int ocfs2_check_range_for_holes(struct inode *inode, loff_t pos,
1100 unsigned int extent_flags;
1101 u32 cpos, clusters, extent_len, phys_cpos;
1102 struct super_block *sb = inode->i_sb;
1104 cpos = pos >> OCFS2_SB(sb)->s_clustersize_bits;
1105 clusters = ocfs2_clusters_for_bytes(sb, pos + count) - cpos;
1108 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos, &extent_len,
1115 if (phys_cpos == 0 || (extent_flags & OCFS2_EXT_UNWRITTEN)) {
1120 if (extent_len > clusters)
1121 extent_len = clusters;
1123 clusters -= extent_len;
1130 static int ocfs2_prepare_inode_for_write(struct dentry *dentry,
1136 int ret = 0, meta_level = appending;
1137 struct inode *inode = dentry->d_inode;
1139 loff_t newsize, saved_pos;
1142 * We sample i_size under a read level meta lock to see if our write
1143 * is extending the file, if it is we back off and get a write level
1147 ret = ocfs2_meta_lock(inode, NULL, meta_level);
1154 /* Clear suid / sgid if necessary. We do this here
1155 * instead of later in the write path because
1156 * remove_suid() calls ->setattr without any hint that
1157 * we may have already done our cluster locking. Since
1158 * ocfs2_setattr() *must* take cluster locks to
1159 * proceeed, this will lead us to recursively lock the
1160 * inode. There's also the dinode i_size state which
1161 * can be lost via setattr during extending writes (we
1162 * set inode->i_size at the end of a write. */
1163 if (should_remove_suid(dentry)) {
1164 if (meta_level == 0) {
1165 ocfs2_meta_unlock(inode, meta_level);
1170 ret = ocfs2_write_remove_suid(inode);
1177 /* work on a copy of ppos until we're sure that we won't have
1178 * to recalculate it due to relocking. */
1180 saved_pos = i_size_read(inode);
1181 mlog(0, "O_APPEND: inode->i_size=%llu\n", saved_pos);
1186 if (ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb))) {
1187 loff_t end = saved_pos + count;
1190 * Skip the O_DIRECT checks if we don't need
1193 if (!direct_io || !(*direct_io))
1197 * Allowing concurrent direct writes means
1198 * i_size changes wouldn't be synchronized, so
1199 * one node could wind up truncating another
1202 if (end > i_size_read(inode)) {
1208 * We don't fill holes during direct io, so
1209 * check for them here. If any are found, the
1210 * caller will have to retake some cluster
1211 * locks and initiate the io as buffered.
1213 ret = ocfs2_check_range_for_holes(inode, saved_pos,
1224 * The rest of this loop is concerned with legacy file
1225 * systems which don't support sparse files.
1228 newsize = count + saved_pos;
1230 mlog(0, "pos=%lld newsize=%lld cursize=%lld\n",
1231 (long long) saved_pos, (long long) newsize,
1232 (long long) i_size_read(inode));
1234 /* No need for a higher level metadata lock if we're
1235 * never going past i_size. */
1236 if (newsize <= i_size_read(inode))
1239 if (meta_level == 0) {
1240 ocfs2_meta_unlock(inode, meta_level);
1245 spin_lock(&OCFS2_I(inode)->ip_lock);
1246 clusters = ocfs2_clusters_for_bytes(inode->i_sb, newsize) -
1247 OCFS2_I(inode)->ip_clusters;
1248 spin_unlock(&OCFS2_I(inode)->ip_lock);
1250 mlog(0, "Writing at EOF, may need more allocation: "
1251 "i_size = %lld, newsize = %lld, need %u clusters\n",
1252 (long long) i_size_read(inode), (long long) newsize,
1255 /* We only want to continue the rest of this loop if
1256 * our extend will actually require more
1261 ret = ocfs2_extend_file(inode, NULL, newsize, count);
1274 ocfs2_meta_unlock(inode, meta_level);
1281 ocfs2_set_next_iovec(const struct iovec **iovp, size_t *basep, size_t bytes)
1283 const struct iovec *iov = *iovp;
1284 size_t base = *basep;
1287 int copy = min(bytes, iov->iov_len - base);
1291 if (iov->iov_len == base) {
1300 static struct page * ocfs2_get_write_source(struct ocfs2_buffered_write_priv *bp,
1301 const struct iovec *cur_iov,
1306 struct page *src_page = NULL;
1308 buf = cur_iov->iov_base + iov_offset;
1310 if (!segment_eq(get_fs(), KERNEL_DS)) {
1312 * Pull in the user page. We want to do this outside
1313 * of the meta data locks in order to preserve locking
1314 * order in case of page fault.
1316 ret = get_user_pages(current, current->mm,
1317 (unsigned long)buf & PAGE_CACHE_MASK, 1,
1318 0, 0, &src_page, NULL);
1320 bp->b_src_buf = kmap(src_page);
1322 src_page = ERR_PTR(-EFAULT);
1324 bp->b_src_buf = buf;
1330 static void ocfs2_put_write_source(struct ocfs2_buffered_write_priv *bp,
1335 page_cache_release(page);
1339 static ssize_t ocfs2_file_buffered_write(struct file *file, loff_t *ppos,
1340 const struct iovec *iov,
1341 unsigned long nr_segs,
1343 ssize_t o_direct_written)
1346 ssize_t copied, total = 0;
1347 size_t iov_offset = 0;
1348 const struct iovec *cur_iov = iov;
1349 struct ocfs2_buffered_write_priv bp;
1353 * handle partial DIO write. Adjust cur_iov if needed.
1355 ocfs2_set_next_iovec(&cur_iov, &iov_offset, o_direct_written);
1358 bp.b_cur_off = iov_offset;
1359 bp.b_cur_iov = cur_iov;
1361 page = ocfs2_get_write_source(&bp, cur_iov, iov_offset);
1363 ret = PTR_ERR(page);
1367 copied = ocfs2_buffered_write_cluster(file, *ppos, count,
1368 ocfs2_map_and_write_user_data,
1371 ocfs2_put_write_source(&bp, page);
1380 *ppos = *ppos + copied;
1383 ocfs2_set_next_iovec(&cur_iov, &iov_offset, copied);
1387 return total ? total : ret;
1390 static int ocfs2_check_iovec(const struct iovec *iov, size_t *counted,
1391 unsigned long *nr_segs)
1393 size_t ocount; /* original count */
1397 for (seg = 0; seg < *nr_segs; seg++) {
1398 const struct iovec *iv = &iov[seg];
1401 * If any segment has a negative length, or the cumulative
1402 * length ever wraps negative then return -EINVAL.
1404 ocount += iv->iov_len;
1405 if (unlikely((ssize_t)(ocount|iv->iov_len) < 0))
1407 if (access_ok(VERIFY_READ, iv->iov_base, iv->iov_len))
1412 ocount -= iv->iov_len; /* This segment is no good */
1420 static ssize_t ocfs2_file_aio_write(struct kiocb *iocb,
1421 const struct iovec *iov,
1422 unsigned long nr_segs,
1425 int ret, direct_io, appending, rw_level, have_alloc_sem = 0;
1426 int can_do_direct, sync = 0;
1427 ssize_t written = 0;
1428 size_t ocount; /* original count */
1429 size_t count; /* after file limit checks */
1430 loff_t *ppos = &iocb->ki_pos;
1431 struct file *file = iocb->ki_filp;
1432 struct inode *inode = file->f_path.dentry->d_inode;
1434 mlog_entry("(0x%p, %u, '%.*s')\n", file,
1435 (unsigned int)nr_segs,
1436 file->f_path.dentry->d_name.len,
1437 file->f_path.dentry->d_name.name);
1439 if (iocb->ki_left == 0)
1442 ret = ocfs2_check_iovec(iov, &ocount, &nr_segs);
1448 vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
1450 appending = file->f_flags & O_APPEND ? 1 : 0;
1451 direct_io = file->f_flags & O_DIRECT ? 1 : 0;
1453 mutex_lock(&inode->i_mutex);
1456 /* to match setattr's i_mutex -> i_alloc_sem -> rw_lock ordering */
1458 down_read(&inode->i_alloc_sem);
1462 /* concurrent O_DIRECT writes are allowed */
1463 rw_level = !direct_io;
1464 ret = ocfs2_rw_lock(inode, rw_level);
1470 can_do_direct = direct_io;
1471 ret = ocfs2_prepare_inode_for_write(file->f_path.dentry, ppos,
1472 iocb->ki_left, appending,
1480 * We can't complete the direct I/O as requested, fall back to
1483 if (direct_io && !can_do_direct) {
1484 ocfs2_rw_unlock(inode, rw_level);
1485 up_read(&inode->i_alloc_sem);
1495 if (!sync && ((file->f_flags & O_SYNC) || IS_SYNC(inode)))
1499 * XXX: Is it ok to execute these checks a second time?
1501 ret = generic_write_checks(file, ppos, &count, S_ISBLK(inode->i_mode));
1506 * Set pos so that sync_page_range_nolock() below understands
1507 * where to start from. We might've moved it around via the
1508 * calls above. The range we want to actually sync starts from
1514 /* communicate with ocfs2_dio_end_io */
1515 ocfs2_iocb_set_rw_locked(iocb);
1518 written = generic_file_direct_write(iocb, iov, &nr_segs, *ppos,
1519 ppos, count, ocount);
1525 written = ocfs2_file_buffered_write(file, ppos, iov, nr_segs,
1529 if (ret != -EFAULT || ret != -ENOSPC)
1536 /* buffered aio wouldn't have proper lock coverage today */
1537 BUG_ON(ret == -EIOCBQUEUED && !(file->f_flags & O_DIRECT));
1540 * deep in g_f_a_w_n()->ocfs2_direct_IO we pass in a ocfs2_dio_end_io
1541 * function pointer which is called when o_direct io completes so that
1542 * it can unlock our rw lock. (it's the clustered equivalent of
1543 * i_alloc_sem; protects truncate from racing with pending ios).
1544 * Unfortunately there are error cases which call end_io and others
1545 * that don't. so we don't have to unlock the rw_lock if either an
1546 * async dio is going to do it in the future or an end_io after an
1547 * error has already done it.
1549 if (ret == -EIOCBQUEUED || !ocfs2_iocb_is_rw_locked(iocb)) {
1556 ocfs2_rw_unlock(inode, rw_level);
1560 up_read(&inode->i_alloc_sem);
1562 if (written > 0 && sync) {
1565 err = sync_page_range_nolock(inode, file->f_mapping, pos, count);
1570 mutex_unlock(&inode->i_mutex);
1573 return written ? written : ret;
1576 static ssize_t ocfs2_file_splice_write(struct pipe_inode_info *pipe,
1583 struct inode *inode = out->f_path.dentry->d_inode;
1585 mlog_entry("(0x%p, 0x%p, %u, '%.*s')\n", out, pipe,
1587 out->f_path.dentry->d_name.len,
1588 out->f_path.dentry->d_name.name);
1590 inode_double_lock(inode, pipe->inode);
1592 ret = ocfs2_rw_lock(inode, 1);
1598 ret = ocfs2_prepare_inode_for_write(out->f_path.dentry, ppos, len, 0,
1605 /* ok, we're done with i_size and alloc work */
1606 ret = generic_file_splice_write_nolock(pipe, out, ppos, len, flags);
1609 ocfs2_rw_unlock(inode, 1);
1611 inode_double_unlock(inode, pipe->inode);
1617 static ssize_t ocfs2_file_splice_read(struct file *in,
1619 struct pipe_inode_info *pipe,
1624 struct inode *inode = in->f_path.dentry->d_inode;
1626 mlog_entry("(0x%p, 0x%p, %u, '%.*s')\n", in, pipe,
1628 in->f_path.dentry->d_name.len,
1629 in->f_path.dentry->d_name.name);
1632 * See the comment in ocfs2_file_aio_read()
1634 ret = ocfs2_meta_lock(inode, NULL, 0);
1639 ocfs2_meta_unlock(inode, 0);
1641 ret = generic_file_splice_read(in, ppos, pipe, len, flags);
1648 static ssize_t ocfs2_file_aio_read(struct kiocb *iocb,
1649 const struct iovec *iov,
1650 unsigned long nr_segs,
1653 int ret = 0, rw_level = -1, have_alloc_sem = 0, lock_level = 0;
1654 struct file *filp = iocb->ki_filp;
1655 struct inode *inode = filp->f_path.dentry->d_inode;
1657 mlog_entry("(0x%p, %u, '%.*s')\n", filp,
1658 (unsigned int)nr_segs,
1659 filp->f_path.dentry->d_name.len,
1660 filp->f_path.dentry->d_name.name);
1669 * buffered reads protect themselves in ->readpage(). O_DIRECT reads
1670 * need locks to protect pending reads from racing with truncate.
1672 if (filp->f_flags & O_DIRECT) {
1673 down_read(&inode->i_alloc_sem);
1676 ret = ocfs2_rw_lock(inode, 0);
1682 /* communicate with ocfs2_dio_end_io */
1683 ocfs2_iocb_set_rw_locked(iocb);
1687 * We're fine letting folks race truncates and extending
1688 * writes with read across the cluster, just like they can
1689 * locally. Hence no rw_lock during read.
1691 * Take and drop the meta data lock to update inode fields
1692 * like i_size. This allows the checks down below
1693 * generic_file_aio_read() a chance of actually working.
1695 ret = ocfs2_meta_lock_atime(inode, filp->f_vfsmnt, &lock_level);
1700 ocfs2_meta_unlock(inode, lock_level);
1702 ret = generic_file_aio_read(iocb, iov, nr_segs, iocb->ki_pos);
1704 mlog(ML_ERROR, "generic_file_aio_read returned -EINVAL\n");
1706 /* buffered aio wouldn't have proper lock coverage today */
1707 BUG_ON(ret == -EIOCBQUEUED && !(filp->f_flags & O_DIRECT));
1709 /* see ocfs2_file_aio_write */
1710 if (ret == -EIOCBQUEUED || !ocfs2_iocb_is_rw_locked(iocb)) {
1717 up_read(&inode->i_alloc_sem);
1719 ocfs2_rw_unlock(inode, rw_level);
1725 const struct inode_operations ocfs2_file_iops = {
1726 .setattr = ocfs2_setattr,
1727 .getattr = ocfs2_getattr,
1728 .permission = ocfs2_permission,
1731 const struct inode_operations ocfs2_special_file_iops = {
1732 .setattr = ocfs2_setattr,
1733 .getattr = ocfs2_getattr,
1734 .permission = ocfs2_permission,
1737 const struct file_operations ocfs2_fops = {
1738 .read = do_sync_read,
1739 .write = do_sync_write,
1740 .sendfile = generic_file_sendfile,
1742 .fsync = ocfs2_sync_file,
1743 .release = ocfs2_file_release,
1744 .open = ocfs2_file_open,
1745 .aio_read = ocfs2_file_aio_read,
1746 .aio_write = ocfs2_file_aio_write,
1747 .ioctl = ocfs2_ioctl,
1748 .splice_read = ocfs2_file_splice_read,
1749 .splice_write = ocfs2_file_splice_write,
1752 const struct file_operations ocfs2_dops = {
1753 .read = generic_read_dir,
1754 .readdir = ocfs2_readdir,
1755 .fsync = ocfs2_sync_file,
1756 .ioctl = ocfs2_ioctl,