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>
37 #define MLOG_MASK_PREFIX ML_INODE
38 #include <cluster/masklog.h>
46 #include "extent_map.h"
56 #include "buffer_head_io.h"
58 static int ocfs2_sync_inode(struct inode *inode)
60 filemap_fdatawrite(inode->i_mapping);
61 return sync_mapping_buffers(inode->i_mapping);
64 static int ocfs2_file_open(struct inode *inode, struct file *file)
67 int mode = file->f_flags;
68 struct ocfs2_inode_info *oi = OCFS2_I(inode);
70 mlog_entry("(0x%p, 0x%p, '%.*s')\n", inode, file,
71 file->f_path.dentry->d_name.len, file->f_path.dentry->d_name.name);
73 spin_lock(&oi->ip_lock);
75 /* Check that the inode hasn't been wiped from disk by another
76 * node. If it hasn't then we're safe as long as we hold the
77 * spin lock until our increment of open count. */
78 if (OCFS2_I(inode)->ip_flags & OCFS2_INODE_DELETED) {
79 spin_unlock(&oi->ip_lock);
86 oi->ip_flags |= OCFS2_INODE_OPEN_DIRECT;
89 spin_unlock(&oi->ip_lock);
96 static int ocfs2_file_release(struct inode *inode, struct file *file)
98 struct ocfs2_inode_info *oi = OCFS2_I(inode);
100 mlog_entry("(0x%p, 0x%p, '%.*s')\n", inode, file,
101 file->f_path.dentry->d_name.len,
102 file->f_path.dentry->d_name.name);
104 spin_lock(&oi->ip_lock);
105 if (!--oi->ip_open_count)
106 oi->ip_flags &= ~OCFS2_INODE_OPEN_DIRECT;
107 spin_unlock(&oi->ip_lock);
114 static int ocfs2_sync_file(struct file *file,
115 struct dentry *dentry,
120 struct inode *inode = dentry->d_inode;
121 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
123 mlog_entry("(0x%p, 0x%p, %d, '%.*s')\n", file, dentry, datasync,
124 dentry->d_name.len, dentry->d_name.name);
126 err = ocfs2_sync_inode(dentry->d_inode);
130 journal = osb->journal->j_journal;
131 err = journal_force_commit(journal);
136 return (err < 0) ? -EIO : 0;
139 int ocfs2_should_update_atime(struct inode *inode,
140 struct vfsmount *vfsmnt)
143 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
145 if (ocfs2_is_hard_readonly(osb) || ocfs2_is_soft_readonly(osb))
148 if ((inode->i_flags & S_NOATIME) ||
149 ((inode->i_sb->s_flags & MS_NODIRATIME) && S_ISDIR(inode->i_mode)))
152 if ((vfsmnt->mnt_flags & MNT_NOATIME) ||
153 ((vfsmnt->mnt_flags & MNT_NODIRATIME) && S_ISDIR(inode->i_mode)))
156 if (vfsmnt->mnt_flags & MNT_RELATIME) {
157 if ((timespec_compare(&inode->i_atime, &inode->i_mtime) <= 0) ||
158 (timespec_compare(&inode->i_atime, &inode->i_ctime) <= 0))
165 if ((now.tv_sec - inode->i_atime.tv_sec <= osb->s_atime_quantum))
171 int ocfs2_update_inode_atime(struct inode *inode,
172 struct buffer_head *bh)
175 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
180 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
181 if (handle == NULL) {
187 inode->i_atime = CURRENT_TIME;
188 ret = ocfs2_mark_inode_dirty(handle, inode, bh);
192 ocfs2_commit_trans(OCFS2_SB(inode->i_sb), handle);
198 int ocfs2_set_inode_size(handle_t *handle,
200 struct buffer_head *fe_bh,
206 i_size_write(inode, new_i_size);
207 inode->i_blocks = ocfs2_align_bytes_to_sectors(new_i_size);
208 inode->i_ctime = inode->i_mtime = CURRENT_TIME;
210 status = ocfs2_mark_inode_dirty(handle, inode, fe_bh);
221 static int ocfs2_simple_size_update(struct inode *inode,
222 struct buffer_head *di_bh,
226 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
227 handle_t *handle = NULL;
229 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
230 if (handle == NULL) {
236 ret = ocfs2_set_inode_size(handle, inode, di_bh,
241 ocfs2_commit_trans(osb, handle);
246 static int ocfs2_orphan_for_truncate(struct ocfs2_super *osb,
248 struct buffer_head *fe_bh,
256 /* TODO: This needs to actually orphan the inode in this
259 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
260 if (IS_ERR(handle)) {
261 status = PTR_ERR(handle);
266 status = ocfs2_set_inode_size(handle, inode, fe_bh, new_i_size);
270 ocfs2_commit_trans(osb, handle);
276 static int ocfs2_truncate_file(struct inode *inode,
277 struct buffer_head *di_bh,
281 struct ocfs2_dinode *fe = NULL;
282 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
283 struct ocfs2_truncate_context *tc = NULL;
285 mlog_entry("(inode = %llu, new_i_size = %llu\n",
286 (unsigned long long)OCFS2_I(inode)->ip_blkno,
287 (unsigned long long)new_i_size);
289 truncate_inode_pages(inode->i_mapping, new_i_size);
291 fe = (struct ocfs2_dinode *) di_bh->b_data;
292 if (!OCFS2_IS_VALID_DINODE(fe)) {
293 OCFS2_RO_ON_INVALID_DINODE(inode->i_sb, fe);
298 mlog_bug_on_msg(le64_to_cpu(fe->i_size) != i_size_read(inode),
299 "Inode %llu, inode i_size = %lld != di "
300 "i_size = %llu, i_flags = 0x%x\n",
301 (unsigned long long)OCFS2_I(inode)->ip_blkno,
303 (unsigned long long)le64_to_cpu(fe->i_size),
304 le32_to_cpu(fe->i_flags));
306 if (new_i_size > le64_to_cpu(fe->i_size)) {
307 mlog(0, "asked to truncate file with size (%llu) to size (%llu)!\n",
308 (unsigned long long)le64_to_cpu(fe->i_size),
309 (unsigned long long)new_i_size);
315 mlog(0, "inode %llu, i_size = %llu, new_i_size = %llu\n",
316 (unsigned long long)le64_to_cpu(fe->i_blkno),
317 (unsigned long long)le64_to_cpu(fe->i_size),
318 (unsigned long long)new_i_size);
320 /* lets handle the simple truncate cases before doing any more
321 * cluster locking. */
322 if (new_i_size == le64_to_cpu(fe->i_size))
325 /* This forces other nodes to sync and drop their pages. Do
326 * this even if we have a truncate without allocation change -
327 * ocfs2 cluster sizes can be much greater than page size, so
328 * we have to truncate them anyway. */
329 status = ocfs2_data_lock(inode, 1);
334 ocfs2_data_unlock(inode, 1);
336 if (le32_to_cpu(fe->i_clusters) ==
337 ocfs2_clusters_for_bytes(osb->sb, new_i_size)) {
338 mlog(0, "fe->i_clusters = %u, so we do a simple truncate\n",
340 /* No allocation change is required, so lets fast path
342 status = ocfs2_simple_size_update(inode, di_bh, new_i_size);
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,
390 struct buffer_head *fe_bh,
392 struct ocfs2_alloc_context *data_ac,
393 struct ocfs2_alloc_context *meta_ac,
394 enum ocfs2_alloc_restarted *reason_ret)
398 struct ocfs2_dinode *fe = (struct ocfs2_dinode *) fe_bh->b_data;
399 enum ocfs2_alloc_restarted reason = RESTART_NONE;
400 u32 bit_off, num_bits;
403 BUG_ON(!clusters_to_add);
405 free_extents = ocfs2_num_free_extents(osb, inode, fe);
406 if (free_extents < 0) {
407 status = free_extents;
412 /* there are two cases which could cause us to EAGAIN in the
413 * we-need-more-metadata case:
414 * 1) we haven't reserved *any*
415 * 2) we are so fragmented, we've needed to add metadata too
417 if (!free_extents && !meta_ac) {
418 mlog(0, "we haven't reserved any metadata!\n");
420 reason = RESTART_META;
422 } else if ((!free_extents)
423 && (ocfs2_alloc_context_bits_left(meta_ac)
424 < ocfs2_extend_meta_needed(fe))) {
425 mlog(0, "filesystem is really fragmented...\n");
427 reason = RESTART_META;
431 status = ocfs2_claim_clusters(osb, handle, data_ac, 1,
432 &bit_off, &num_bits);
434 if (status != -ENOSPC)
439 BUG_ON(num_bits > clusters_to_add);
441 /* reserve our write early -- insert_extent may update the inode */
442 status = ocfs2_journal_access(handle, inode, fe_bh,
443 OCFS2_JOURNAL_ACCESS_WRITE);
449 block = ocfs2_clusters_to_blocks(osb->sb, bit_off);
450 mlog(0, "Allocating %u clusters at block %u for inode %llu\n",
451 num_bits, bit_off, (unsigned long long)OCFS2_I(inode)->ip_blkno);
452 status = ocfs2_insert_extent(osb, handle, inode, fe_bh, block,
459 le32_add_cpu(&fe->i_clusters, num_bits);
460 spin_lock(&OCFS2_I(inode)->ip_lock);
461 OCFS2_I(inode)->ip_clusters = le32_to_cpu(fe->i_clusters);
462 spin_unlock(&OCFS2_I(inode)->ip_lock);
464 status = ocfs2_journal_dirty(handle, fe_bh);
470 clusters_to_add -= num_bits;
472 if (clusters_to_add) {
473 mlog(0, "need to alloc once more, clusters = %u, wanted = "
474 "%u\n", fe->i_clusters, clusters_to_add);
476 reason = RESTART_TRANS;
482 *reason_ret = reason;
486 static int ocfs2_extend_allocation(struct inode *inode,
490 int restart_func = 0;
491 int drop_alloc_sem = 0;
492 int credits, num_free_extents;
494 struct buffer_head *bh = NULL;
495 struct ocfs2_dinode *fe = NULL;
496 handle_t *handle = NULL;
497 struct ocfs2_alloc_context *data_ac = NULL;
498 struct ocfs2_alloc_context *meta_ac = NULL;
499 enum ocfs2_alloc_restarted why;
500 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
502 mlog_entry("(clusters_to_add = %u)\n", clusters_to_add);
504 status = ocfs2_read_block(osb, OCFS2_I(inode)->ip_blkno, &bh,
505 OCFS2_BH_CACHED, inode);
511 fe = (struct ocfs2_dinode *) bh->b_data;
512 if (!OCFS2_IS_VALID_DINODE(fe)) {
513 OCFS2_RO_ON_INVALID_DINODE(inode->i_sb, fe);
519 BUG_ON(le32_to_cpu(fe->i_clusters) != OCFS2_I(inode)->ip_clusters);
521 mlog(0, "extend inode %llu, i_size = %lld, fe->i_clusters = %u, "
522 "clusters_to_add = %u\n",
523 (unsigned long long)OCFS2_I(inode)->ip_blkno, i_size_read(inode),
524 fe->i_clusters, clusters_to_add);
526 num_free_extents = ocfs2_num_free_extents(osb,
529 if (num_free_extents < 0) {
530 status = num_free_extents;
535 if (!num_free_extents) {
536 status = ocfs2_reserve_new_metadata(osb, fe, &meta_ac);
538 if (status != -ENOSPC)
544 status = ocfs2_reserve_clusters(osb, clusters_to_add, &data_ac);
546 if (status != -ENOSPC)
551 /* blocks peope in read/write from reading our allocation
552 * until we're done changing it. We depend on i_mutex to block
553 * other extend/truncate calls while we're here. Ordering wrt
554 * start_trans is important here -- always do it before! */
555 down_write(&OCFS2_I(inode)->ip_alloc_sem);
558 credits = ocfs2_calc_extend_credits(osb->sb, fe, clusters_to_add);
559 handle = ocfs2_start_trans(osb, credits);
560 if (IS_ERR(handle)) {
561 status = PTR_ERR(handle);
567 restarted_transaction:
568 /* reserve a write to the file entry early on - that we if we
569 * run out of credits in the allocation path, we can still
571 status = ocfs2_journal_access(handle, inode, bh,
572 OCFS2_JOURNAL_ACCESS_WRITE);
578 prev_clusters = OCFS2_I(inode)->ip_clusters;
580 status = ocfs2_do_extend_allocation(osb,
588 if ((status < 0) && (status != -EAGAIN)) {
589 if (status != -ENOSPC)
594 status = ocfs2_journal_dirty(handle, bh);
600 spin_lock(&OCFS2_I(inode)->ip_lock);
601 clusters_to_add -= (OCFS2_I(inode)->ip_clusters - prev_clusters);
602 spin_unlock(&OCFS2_I(inode)->ip_lock);
604 if (why != RESTART_NONE && clusters_to_add) {
605 if (why == RESTART_META) {
606 mlog(0, "restarting function.\n");
609 BUG_ON(why != RESTART_TRANS);
611 mlog(0, "restarting transaction.\n");
612 /* TODO: This can be more intelligent. */
613 credits = ocfs2_calc_extend_credits(osb->sb,
616 status = ocfs2_extend_trans(handle, credits);
618 /* handle still has to be committed at
624 goto restarted_transaction;
628 mlog(0, "fe: i_clusters = %u, i_size=%llu\n",
629 fe->i_clusters, (unsigned long long)fe->i_size);
630 mlog(0, "inode: ip_clusters=%u, i_size=%lld\n",
631 OCFS2_I(inode)->ip_clusters, i_size_read(inode));
634 if (drop_alloc_sem) {
635 up_write(&OCFS2_I(inode)->ip_alloc_sem);
639 ocfs2_commit_trans(osb, handle);
643 ocfs2_free_alloc_context(data_ac);
647 ocfs2_free_alloc_context(meta_ac);
650 if ((!status) && restart_func) {
663 /* Some parts of this taken from generic_cont_expand, which turned out
664 * to be too fragile to do exactly what we need without us having to
665 * worry about recursive locking in ->prepare_write() and
666 * ->commit_write(). */
667 static int ocfs2_write_zero_page(struct inode *inode,
670 struct address_space *mapping = inode->i_mapping;
674 handle_t *handle = NULL;
677 offset = (size & (PAGE_CACHE_SIZE-1)); /* Within page */
678 /* ugh. in prepare/commit_write, if from==to==start of block, we
679 ** skip the prepare. make sure we never send an offset for the start
682 if ((offset & (inode->i_sb->s_blocksize - 1)) == 0) {
685 index = size >> PAGE_CACHE_SHIFT;
687 page = grab_cache_page(mapping, index);
694 ret = ocfs2_prepare_write_nolock(inode, page, offset, offset);
700 if (ocfs2_should_order_data(inode)) {
701 handle = ocfs2_start_walk_page_trans(inode, page, offset,
703 if (IS_ERR(handle)) {
704 ret = PTR_ERR(handle);
710 /* must not update i_size! */
711 ret = block_commit_write(page, offset, offset);
718 ocfs2_commit_trans(OCFS2_SB(inode->i_sb), handle);
721 page_cache_release(page);
726 static int ocfs2_zero_extend(struct inode *inode,
731 struct super_block *sb = inode->i_sb;
733 start_off = ocfs2_align_bytes_to_blocks(sb, i_size_read(inode));
734 while (start_off < zero_to_size) {
735 ret = ocfs2_write_zero_page(inode, start_off);
741 start_off += sb->s_blocksize;
744 * Very large extends have the potential to lock up
745 * the cpu for extended periods of time.
755 * A tail_to_skip value > 0 indicates that we're being called from
756 * ocfs2_file_aio_write(). This has the following implications:
758 * - we don't want to update i_size
759 * - di_bh will be NULL, which is fine because it's only used in the
760 * case where we want to update i_size.
761 * - ocfs2_zero_extend() will then only be filling the hole created
762 * between i_size and the start of the write.
764 static int ocfs2_extend_file(struct inode *inode,
765 struct buffer_head *di_bh,
772 BUG_ON(!tail_to_skip && !di_bh);
774 /* setattr sometimes calls us like this. */
778 if (i_size_read(inode) == new_i_size)
780 BUG_ON(new_i_size < i_size_read(inode));
782 clusters_to_add = ocfs2_clusters_for_bytes(inode->i_sb, new_i_size) -
783 OCFS2_I(inode)->ip_clusters;
786 * protect the pages that ocfs2_zero_extend is going to be
787 * pulling into the page cache.. we do this before the
788 * metadata extend so that we don't get into the situation
789 * where we've extended the metadata but can't get the data
792 ret = ocfs2_data_lock(inode, 1);
798 if (clusters_to_add) {
799 ret = ocfs2_extend_allocation(inode, clusters_to_add);
807 * Call this even if we don't add any clusters to the tree. We
808 * still need to zero the area between the old i_size and the
811 ret = ocfs2_zero_extend(inode, (u64)new_i_size - tail_to_skip);
818 /* We're being called from ocfs2_setattr() which wants
819 * us to update i_size */
820 ret = ocfs2_simple_size_update(inode, di_bh, new_i_size);
826 ocfs2_data_unlock(inode, 1);
832 int ocfs2_setattr(struct dentry *dentry, struct iattr *attr)
834 int status = 0, size_change;
835 struct inode *inode = dentry->d_inode;
836 struct super_block *sb = inode->i_sb;
837 struct ocfs2_super *osb = OCFS2_SB(sb);
838 struct buffer_head *bh = NULL;
839 handle_t *handle = NULL;
841 mlog_entry("(0x%p, '%.*s')\n", dentry,
842 dentry->d_name.len, dentry->d_name.name);
844 if (attr->ia_valid & ATTR_MODE)
845 mlog(0, "mode change: %d\n", attr->ia_mode);
846 if (attr->ia_valid & ATTR_UID)
847 mlog(0, "uid change: %d\n", attr->ia_uid);
848 if (attr->ia_valid & ATTR_GID)
849 mlog(0, "gid change: %d\n", attr->ia_gid);
850 if (attr->ia_valid & ATTR_SIZE)
851 mlog(0, "size change...\n");
852 if (attr->ia_valid & (ATTR_ATIME | ATTR_MTIME | ATTR_CTIME))
853 mlog(0, "time change...\n");
855 #define OCFS2_VALID_ATTRS (ATTR_ATIME | ATTR_MTIME | ATTR_CTIME | ATTR_SIZE \
856 | ATTR_GID | ATTR_UID | ATTR_MODE)
857 if (!(attr->ia_valid & OCFS2_VALID_ATTRS)) {
858 mlog(0, "can't handle attrs: 0x%x\n", attr->ia_valid);
862 status = inode_change_ok(inode, attr);
866 size_change = S_ISREG(inode->i_mode) && attr->ia_valid & ATTR_SIZE;
868 status = ocfs2_rw_lock(inode, 1);
875 status = ocfs2_meta_lock(inode, &bh, 1);
877 if (status != -ENOENT)
882 if (size_change && attr->ia_size != i_size_read(inode)) {
883 if (i_size_read(inode) > attr->ia_size)
884 status = ocfs2_truncate_file(inode, bh, attr->ia_size);
886 status = ocfs2_extend_file(inode, bh, attr->ia_size, 0);
888 if (status != -ENOSPC)
895 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
896 if (IS_ERR(handle)) {
897 status = PTR_ERR(handle);
902 status = inode_setattr(inode, attr);
908 status = ocfs2_mark_inode_dirty(handle, inode, bh);
913 ocfs2_commit_trans(osb, handle);
915 ocfs2_meta_unlock(inode, 1);
918 ocfs2_rw_unlock(inode, 1);
927 int ocfs2_getattr(struct vfsmount *mnt,
928 struct dentry *dentry,
931 struct inode *inode = dentry->d_inode;
932 struct super_block *sb = dentry->d_inode->i_sb;
933 struct ocfs2_super *osb = sb->s_fs_info;
938 err = ocfs2_inode_revalidate(dentry);
945 generic_fillattr(inode, stat);
947 /* We set the blksize from the cluster size for performance */
948 stat->blksize = osb->s_clustersize;
956 int ocfs2_permission(struct inode *inode, int mask, struct nameidata *nd)
962 ret = ocfs2_meta_lock(inode, NULL, 0);
968 ret = generic_permission(inode, mask, NULL);
970 ocfs2_meta_unlock(inode, 0);
976 static int ocfs2_write_remove_suid(struct inode *inode)
979 struct buffer_head *bh = NULL;
980 struct ocfs2_inode_info *oi = OCFS2_I(inode);
982 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
983 struct ocfs2_dinode *di;
985 mlog_entry("(Inode %llu, mode 0%o)\n",
986 (unsigned long long)oi->ip_blkno, inode->i_mode);
988 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
989 if (handle == NULL) {
995 ret = ocfs2_read_block(osb, oi->ip_blkno, &bh, OCFS2_BH_CACHED, inode);
1001 ret = ocfs2_journal_access(handle, inode, bh,
1002 OCFS2_JOURNAL_ACCESS_WRITE);
1008 inode->i_mode &= ~S_ISUID;
1009 if ((inode->i_mode & S_ISGID) && (inode->i_mode & S_IXGRP))
1010 inode->i_mode &= ~S_ISGID;
1012 di = (struct ocfs2_dinode *) bh->b_data;
1013 di->i_mode = cpu_to_le16(inode->i_mode);
1015 ret = ocfs2_journal_dirty(handle, bh);
1021 ocfs2_commit_trans(osb, handle);
1027 static int ocfs2_prepare_inode_for_write(struct dentry *dentry,
1032 int ret = 0, meta_level = appending;
1033 struct inode *inode = dentry->d_inode;
1035 loff_t newsize, saved_pos;
1038 * We sample i_size under a read level meta lock to see if our write
1039 * is extending the file, if it is we back off and get a write level
1043 ret = ocfs2_meta_lock(inode, NULL, meta_level);
1050 /* Clear suid / sgid if necessary. We do this here
1051 * instead of later in the write path because
1052 * remove_suid() calls ->setattr without any hint that
1053 * we may have already done our cluster locking. Since
1054 * ocfs2_setattr() *must* take cluster locks to
1055 * proceeed, this will lead us to recursively lock the
1056 * inode. There's also the dinode i_size state which
1057 * can be lost via setattr during extending writes (we
1058 * set inode->i_size at the end of a write. */
1059 if (should_remove_suid(dentry)) {
1060 if (meta_level == 0) {
1061 ocfs2_meta_unlock(inode, meta_level);
1066 ret = ocfs2_write_remove_suid(inode);
1073 /* work on a copy of ppos until we're sure that we won't have
1074 * to recalculate it due to relocking. */
1076 saved_pos = i_size_read(inode);
1077 mlog(0, "O_APPEND: inode->i_size=%llu\n", saved_pos);
1081 newsize = count + saved_pos;
1083 mlog(0, "pos=%lld newsize=%lld cursize=%lld\n",
1084 (long long) saved_pos, (long long) newsize,
1085 (long long) i_size_read(inode));
1087 /* No need for a higher level metadata lock if we're
1088 * never going past i_size. */
1089 if (newsize <= i_size_read(inode))
1092 if (meta_level == 0) {
1093 ocfs2_meta_unlock(inode, meta_level);
1098 spin_lock(&OCFS2_I(inode)->ip_lock);
1099 clusters = ocfs2_clusters_for_bytes(inode->i_sb, newsize) -
1100 OCFS2_I(inode)->ip_clusters;
1101 spin_unlock(&OCFS2_I(inode)->ip_lock);
1103 mlog(0, "Writing at EOF, may need more allocation: "
1104 "i_size = %lld, newsize = %lld, need %u clusters\n",
1105 (long long) i_size_read(inode), (long long) newsize,
1108 /* We only want to continue the rest of this loop if
1109 * our extend will actually require more
1114 ret = ocfs2_extend_file(inode, NULL, newsize, count);
1127 ocfs2_meta_unlock(inode, meta_level);
1133 static ssize_t ocfs2_file_aio_write(struct kiocb *iocb,
1134 const struct iovec *iov,
1135 unsigned long nr_segs,
1138 int ret, rw_level, have_alloc_sem = 0;
1139 struct file *filp = iocb->ki_filp;
1140 struct inode *inode = filp->f_path.dentry->d_inode;
1141 int appending = filp->f_flags & O_APPEND ? 1 : 0;
1143 mlog_entry("(0x%p, %u, '%.*s')\n", filp,
1144 (unsigned int)nr_segs,
1145 filp->f_path.dentry->d_name.len,
1146 filp->f_path.dentry->d_name.name);
1148 /* happy write of zero bytes */
1149 if (iocb->ki_left == 0)
1152 mutex_lock(&inode->i_mutex);
1153 /* to match setattr's i_mutex -> i_alloc_sem -> rw_lock ordering */
1154 if (filp->f_flags & O_DIRECT) {
1156 down_read(&inode->i_alloc_sem);
1159 /* concurrent O_DIRECT writes are allowed */
1160 rw_level = (filp->f_flags & O_DIRECT) ? 0 : 1;
1161 ret = ocfs2_rw_lock(inode, rw_level);
1168 ret = ocfs2_prepare_inode_for_write(filp->f_path.dentry, &iocb->ki_pos,
1169 iocb->ki_left, appending);
1175 /* communicate with ocfs2_dio_end_io */
1176 ocfs2_iocb_set_rw_locked(iocb);
1178 ret = generic_file_aio_write_nolock(iocb, iov, nr_segs, iocb->ki_pos);
1180 /* buffered aio wouldn't have proper lock coverage today */
1181 BUG_ON(ret == -EIOCBQUEUED && !(filp->f_flags & O_DIRECT));
1184 * deep in g_f_a_w_n()->ocfs2_direct_IO we pass in a ocfs2_dio_end_io
1185 * function pointer which is called when o_direct io completes so that
1186 * it can unlock our rw lock. (it's the clustered equivalent of
1187 * i_alloc_sem; protects truncate from racing with pending ios).
1188 * Unfortunately there are error cases which call end_io and others
1189 * that don't. so we don't have to unlock the rw_lock if either an
1190 * async dio is going to do it in the future or an end_io after an
1191 * error has already done it.
1193 if (ret == -EIOCBQUEUED || !ocfs2_iocb_is_rw_locked(iocb)) {
1200 up_read(&inode->i_alloc_sem);
1202 ocfs2_rw_unlock(inode, rw_level);
1203 mutex_unlock(&inode->i_mutex);
1209 static ssize_t ocfs2_file_splice_write(struct pipe_inode_info *pipe,
1216 struct inode *inode = out->f_path.dentry->d_inode;
1218 mlog_entry("(0x%p, 0x%p, %u, '%.*s')\n", out, pipe,
1220 out->f_path.dentry->d_name.len,
1221 out->f_path.dentry->d_name.name);
1223 inode_double_lock(inode, pipe->inode);
1225 ret = ocfs2_rw_lock(inode, 1);
1231 ret = ocfs2_prepare_inode_for_write(out->f_path.dentry, ppos, len, 0);
1237 /* ok, we're done with i_size and alloc work */
1238 ret = generic_file_splice_write_nolock(pipe, out, ppos, len, flags);
1241 ocfs2_rw_unlock(inode, 1);
1243 inode_double_unlock(inode, pipe->inode);
1249 static ssize_t ocfs2_file_splice_read(struct file *in,
1251 struct pipe_inode_info *pipe,
1256 struct inode *inode = in->f_path.dentry->d_inode;
1258 mlog_entry("(0x%p, 0x%p, %u, '%.*s')\n", in, pipe,
1260 in->f_path.dentry->d_name.len,
1261 in->f_path.dentry->d_name.name);
1264 * See the comment in ocfs2_file_aio_read()
1266 ret = ocfs2_meta_lock(inode, NULL, 0);
1271 ocfs2_meta_unlock(inode, 0);
1273 ret = generic_file_splice_read(in, ppos, pipe, len, flags);
1280 static ssize_t ocfs2_file_aio_read(struct kiocb *iocb,
1281 const struct iovec *iov,
1282 unsigned long nr_segs,
1285 int ret = 0, rw_level = -1, have_alloc_sem = 0, lock_level = 0;
1286 struct file *filp = iocb->ki_filp;
1287 struct inode *inode = filp->f_path.dentry->d_inode;
1289 mlog_entry("(0x%p, %u, '%.*s')\n", filp,
1290 (unsigned int)nr_segs,
1291 filp->f_path.dentry->d_name.len,
1292 filp->f_path.dentry->d_name.name);
1301 * buffered reads protect themselves in ->readpage(). O_DIRECT reads
1302 * need locks to protect pending reads from racing with truncate.
1304 if (filp->f_flags & O_DIRECT) {
1305 down_read(&inode->i_alloc_sem);
1308 ret = ocfs2_rw_lock(inode, 0);
1314 /* communicate with ocfs2_dio_end_io */
1315 ocfs2_iocb_set_rw_locked(iocb);
1319 * We're fine letting folks race truncates and extending
1320 * writes with read across the cluster, just like they can
1321 * locally. Hence no rw_lock during read.
1323 * Take and drop the meta data lock to update inode fields
1324 * like i_size. This allows the checks down below
1325 * generic_file_aio_read() a chance of actually working.
1327 ret = ocfs2_meta_lock_atime(inode, filp->f_vfsmnt, &lock_level);
1332 ocfs2_meta_unlock(inode, lock_level);
1334 ret = generic_file_aio_read(iocb, iov, nr_segs, iocb->ki_pos);
1336 mlog(ML_ERROR, "generic_file_aio_read returned -EINVAL\n");
1338 /* buffered aio wouldn't have proper lock coverage today */
1339 BUG_ON(ret == -EIOCBQUEUED && !(filp->f_flags & O_DIRECT));
1341 /* see ocfs2_file_aio_write */
1342 if (ret == -EIOCBQUEUED || !ocfs2_iocb_is_rw_locked(iocb)) {
1349 up_read(&inode->i_alloc_sem);
1351 ocfs2_rw_unlock(inode, rw_level);
1357 struct inode_operations ocfs2_file_iops = {
1358 .setattr = ocfs2_setattr,
1359 .getattr = ocfs2_getattr,
1360 .permission = ocfs2_permission,
1363 struct inode_operations ocfs2_special_file_iops = {
1364 .setattr = ocfs2_setattr,
1365 .getattr = ocfs2_getattr,
1366 .permission = ocfs2_permission,
1369 const struct file_operations ocfs2_fops = {
1370 .read = do_sync_read,
1371 .write = do_sync_write,
1372 .sendfile = generic_file_sendfile,
1374 .fsync = ocfs2_sync_file,
1375 .release = ocfs2_file_release,
1376 .open = ocfs2_file_open,
1377 .aio_read = ocfs2_file_aio_read,
1378 .aio_write = ocfs2_file_aio_write,
1379 .ioctl = ocfs2_ioctl,
1380 .splice_read = ocfs2_file_splice_read,
1381 .splice_write = ocfs2_file_splice_write,
1384 const struct file_operations ocfs2_dops = {
1385 .read = generic_read_dir,
1386 .readdir = ocfs2_readdir,
1387 .fsync = ocfs2_sync_file,
1388 .ioctl = ocfs2_ioctl,