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)))
153 * We can be called with no vfsmnt structure - NFSD will
156 * Note that our action here is different than touch_atime() -
157 * if we can't tell whether this is a noatime mount, then we
158 * don't know whether to trust the value of s_atime_quantum.
163 if ((vfsmnt->mnt_flags & MNT_NOATIME) ||
164 ((vfsmnt->mnt_flags & MNT_NODIRATIME) && S_ISDIR(inode->i_mode)))
167 if (vfsmnt->mnt_flags & MNT_RELATIME) {
168 if ((timespec_compare(&inode->i_atime, &inode->i_mtime) <= 0) ||
169 (timespec_compare(&inode->i_atime, &inode->i_ctime) <= 0))
176 if ((now.tv_sec - inode->i_atime.tv_sec <= osb->s_atime_quantum))
182 int ocfs2_update_inode_atime(struct inode *inode,
183 struct buffer_head *bh)
186 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
191 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
192 if (handle == NULL) {
198 inode->i_atime = CURRENT_TIME;
199 ret = ocfs2_mark_inode_dirty(handle, inode, bh);
203 ocfs2_commit_trans(OCFS2_SB(inode->i_sb), handle);
209 int ocfs2_set_inode_size(handle_t *handle,
211 struct buffer_head *fe_bh,
217 i_size_write(inode, new_i_size);
218 inode->i_blocks = ocfs2_align_bytes_to_sectors(new_i_size);
219 inode->i_ctime = inode->i_mtime = CURRENT_TIME;
221 status = ocfs2_mark_inode_dirty(handle, inode, fe_bh);
232 static int ocfs2_simple_size_update(struct inode *inode,
233 struct buffer_head *di_bh,
237 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
238 handle_t *handle = NULL;
240 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
241 if (handle == NULL) {
247 ret = ocfs2_set_inode_size(handle, inode, di_bh,
252 ocfs2_commit_trans(osb, handle);
257 static int ocfs2_orphan_for_truncate(struct ocfs2_super *osb,
259 struct buffer_head *fe_bh,
267 /* TODO: This needs to actually orphan the inode in this
270 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
271 if (IS_ERR(handle)) {
272 status = PTR_ERR(handle);
277 status = ocfs2_set_inode_size(handle, inode, fe_bh, new_i_size);
281 ocfs2_commit_trans(osb, handle);
287 static int ocfs2_truncate_file(struct inode *inode,
288 struct buffer_head *di_bh,
292 struct ocfs2_dinode *fe = NULL;
293 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
294 struct ocfs2_truncate_context *tc = NULL;
296 mlog_entry("(inode = %llu, new_i_size = %llu\n",
297 (unsigned long long)OCFS2_I(inode)->ip_blkno,
298 (unsigned long long)new_i_size);
300 truncate_inode_pages(inode->i_mapping, new_i_size);
302 fe = (struct ocfs2_dinode *) di_bh->b_data;
303 if (!OCFS2_IS_VALID_DINODE(fe)) {
304 OCFS2_RO_ON_INVALID_DINODE(inode->i_sb, fe);
309 mlog_bug_on_msg(le64_to_cpu(fe->i_size) != i_size_read(inode),
310 "Inode %llu, inode i_size = %lld != di "
311 "i_size = %llu, i_flags = 0x%x\n",
312 (unsigned long long)OCFS2_I(inode)->ip_blkno,
314 (unsigned long long)le64_to_cpu(fe->i_size),
315 le32_to_cpu(fe->i_flags));
317 if (new_i_size > le64_to_cpu(fe->i_size)) {
318 mlog(0, "asked to truncate file with size (%llu) to size (%llu)!\n",
319 (unsigned long long)le64_to_cpu(fe->i_size),
320 (unsigned long long)new_i_size);
326 mlog(0, "inode %llu, i_size = %llu, new_i_size = %llu\n",
327 (unsigned long long)le64_to_cpu(fe->i_blkno),
328 (unsigned long long)le64_to_cpu(fe->i_size),
329 (unsigned long long)new_i_size);
331 /* lets handle the simple truncate cases before doing any more
332 * cluster locking. */
333 if (new_i_size == le64_to_cpu(fe->i_size))
336 /* This forces other nodes to sync and drop their pages. Do
337 * this even if we have a truncate without allocation change -
338 * ocfs2 cluster sizes can be much greater than page size, so
339 * we have to truncate them anyway. */
340 status = ocfs2_data_lock(inode, 1);
345 ocfs2_data_unlock(inode, 1);
347 if (le32_to_cpu(fe->i_clusters) ==
348 ocfs2_clusters_for_bytes(osb->sb, new_i_size)) {
349 mlog(0, "fe->i_clusters = %u, so we do a simple truncate\n",
351 /* No allocation change is required, so lets fast path
353 status = ocfs2_simple_size_update(inode, di_bh, new_i_size);
359 /* alright, we're going to need to do a full blown alloc size
360 * change. Orphan the inode so that recovery can complete the
361 * truncate if necessary. This does the task of marking
363 status = ocfs2_orphan_for_truncate(osb, inode, di_bh, new_i_size);
369 status = ocfs2_prepare_truncate(osb, inode, di_bh, &tc);
375 status = ocfs2_commit_truncate(osb, inode, di_bh, tc);
381 /* TODO: orphan dir cleanup here. */
389 * extend allocation only here.
390 * we'll update all the disk stuff, and oip->alloc_size
392 * expect stuff to be locked, a transaction started and enough data /
393 * metadata reservations in the contexts.
395 * Will return -EAGAIN, and a reason if a restart is needed.
396 * If passed in, *reason will always be set, even in error.
398 int ocfs2_do_extend_allocation(struct ocfs2_super *osb,
401 struct buffer_head *fe_bh,
403 struct ocfs2_alloc_context *data_ac,
404 struct ocfs2_alloc_context *meta_ac,
405 enum ocfs2_alloc_restarted *reason_ret)
409 struct ocfs2_dinode *fe = (struct ocfs2_dinode *) fe_bh->b_data;
410 enum ocfs2_alloc_restarted reason = RESTART_NONE;
411 u32 bit_off, num_bits;
414 BUG_ON(!clusters_to_add);
416 free_extents = ocfs2_num_free_extents(osb, inode, fe);
417 if (free_extents < 0) {
418 status = free_extents;
423 /* there are two cases which could cause us to EAGAIN in the
424 * we-need-more-metadata case:
425 * 1) we haven't reserved *any*
426 * 2) we are so fragmented, we've needed to add metadata too
428 if (!free_extents && !meta_ac) {
429 mlog(0, "we haven't reserved any metadata!\n");
431 reason = RESTART_META;
433 } else if ((!free_extents)
434 && (ocfs2_alloc_context_bits_left(meta_ac)
435 < ocfs2_extend_meta_needed(fe))) {
436 mlog(0, "filesystem is really fragmented...\n");
438 reason = RESTART_META;
442 status = ocfs2_claim_clusters(osb, handle, data_ac, 1,
443 &bit_off, &num_bits);
445 if (status != -ENOSPC)
450 BUG_ON(num_bits > clusters_to_add);
452 /* reserve our write early -- insert_extent may update the inode */
453 status = ocfs2_journal_access(handle, inode, fe_bh,
454 OCFS2_JOURNAL_ACCESS_WRITE);
460 block = ocfs2_clusters_to_blocks(osb->sb, bit_off);
461 mlog(0, "Allocating %u clusters at block %u for inode %llu\n",
462 num_bits, bit_off, (unsigned long long)OCFS2_I(inode)->ip_blkno);
463 status = ocfs2_insert_extent(osb, handle, inode, fe_bh, block,
470 le32_add_cpu(&fe->i_clusters, num_bits);
471 spin_lock(&OCFS2_I(inode)->ip_lock);
472 OCFS2_I(inode)->ip_clusters = le32_to_cpu(fe->i_clusters);
473 spin_unlock(&OCFS2_I(inode)->ip_lock);
475 status = ocfs2_journal_dirty(handle, fe_bh);
481 clusters_to_add -= num_bits;
483 if (clusters_to_add) {
484 mlog(0, "need to alloc once more, clusters = %u, wanted = "
485 "%u\n", fe->i_clusters, clusters_to_add);
487 reason = RESTART_TRANS;
493 *reason_ret = reason;
497 static int ocfs2_extend_allocation(struct inode *inode,
501 int restart_func = 0;
502 int drop_alloc_sem = 0;
503 int credits, num_free_extents;
505 struct buffer_head *bh = NULL;
506 struct ocfs2_dinode *fe = NULL;
507 handle_t *handle = NULL;
508 struct ocfs2_alloc_context *data_ac = NULL;
509 struct ocfs2_alloc_context *meta_ac = NULL;
510 enum ocfs2_alloc_restarted why;
511 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
513 mlog_entry("(clusters_to_add = %u)\n", clusters_to_add);
515 status = ocfs2_read_block(osb, OCFS2_I(inode)->ip_blkno, &bh,
516 OCFS2_BH_CACHED, inode);
522 fe = (struct ocfs2_dinode *) bh->b_data;
523 if (!OCFS2_IS_VALID_DINODE(fe)) {
524 OCFS2_RO_ON_INVALID_DINODE(inode->i_sb, fe);
530 BUG_ON(le32_to_cpu(fe->i_clusters) != OCFS2_I(inode)->ip_clusters);
532 mlog(0, "extend inode %llu, i_size = %lld, fe->i_clusters = %u, "
533 "clusters_to_add = %u\n",
534 (unsigned long long)OCFS2_I(inode)->ip_blkno, i_size_read(inode),
535 fe->i_clusters, clusters_to_add);
537 num_free_extents = ocfs2_num_free_extents(osb,
540 if (num_free_extents < 0) {
541 status = num_free_extents;
546 if (!num_free_extents) {
547 status = ocfs2_reserve_new_metadata(osb, fe, &meta_ac);
549 if (status != -ENOSPC)
555 status = ocfs2_reserve_clusters(osb, clusters_to_add, &data_ac);
557 if (status != -ENOSPC)
562 /* blocks peope in read/write from reading our allocation
563 * until we're done changing it. We depend on i_mutex to block
564 * other extend/truncate calls while we're here. Ordering wrt
565 * start_trans is important here -- always do it before! */
566 down_write(&OCFS2_I(inode)->ip_alloc_sem);
569 credits = ocfs2_calc_extend_credits(osb->sb, fe, clusters_to_add);
570 handle = ocfs2_start_trans(osb, credits);
571 if (IS_ERR(handle)) {
572 status = PTR_ERR(handle);
578 restarted_transaction:
579 /* reserve a write to the file entry early on - that we if we
580 * run out of credits in the allocation path, we can still
582 status = ocfs2_journal_access(handle, inode, bh,
583 OCFS2_JOURNAL_ACCESS_WRITE);
589 prev_clusters = OCFS2_I(inode)->ip_clusters;
591 status = ocfs2_do_extend_allocation(osb,
599 if ((status < 0) && (status != -EAGAIN)) {
600 if (status != -ENOSPC)
605 status = ocfs2_journal_dirty(handle, bh);
611 spin_lock(&OCFS2_I(inode)->ip_lock);
612 clusters_to_add -= (OCFS2_I(inode)->ip_clusters - prev_clusters);
613 spin_unlock(&OCFS2_I(inode)->ip_lock);
615 if (why != RESTART_NONE && clusters_to_add) {
616 if (why == RESTART_META) {
617 mlog(0, "restarting function.\n");
620 BUG_ON(why != RESTART_TRANS);
622 mlog(0, "restarting transaction.\n");
623 /* TODO: This can be more intelligent. */
624 credits = ocfs2_calc_extend_credits(osb->sb,
627 status = ocfs2_extend_trans(handle, credits);
629 /* handle still has to be committed at
635 goto restarted_transaction;
639 mlog(0, "fe: i_clusters = %u, i_size=%llu\n",
640 fe->i_clusters, (unsigned long long)fe->i_size);
641 mlog(0, "inode: ip_clusters=%u, i_size=%lld\n",
642 OCFS2_I(inode)->ip_clusters, i_size_read(inode));
645 if (drop_alloc_sem) {
646 up_write(&OCFS2_I(inode)->ip_alloc_sem);
650 ocfs2_commit_trans(osb, handle);
654 ocfs2_free_alloc_context(data_ac);
658 ocfs2_free_alloc_context(meta_ac);
661 if ((!status) && restart_func) {
674 /* Some parts of this taken from generic_cont_expand, which turned out
675 * to be too fragile to do exactly what we need without us having to
676 * worry about recursive locking in ->prepare_write() and
677 * ->commit_write(). */
678 static int ocfs2_write_zero_page(struct inode *inode,
681 struct address_space *mapping = inode->i_mapping;
685 handle_t *handle = NULL;
688 offset = (size & (PAGE_CACHE_SIZE-1)); /* Within page */
689 /* ugh. in prepare/commit_write, if from==to==start of block, we
690 ** skip the prepare. make sure we never send an offset for the start
693 if ((offset & (inode->i_sb->s_blocksize - 1)) == 0) {
696 index = size >> PAGE_CACHE_SHIFT;
698 page = grab_cache_page(mapping, index);
705 ret = ocfs2_prepare_write_nolock(inode, page, offset, offset);
711 if (ocfs2_should_order_data(inode)) {
712 handle = ocfs2_start_walk_page_trans(inode, page, offset,
714 if (IS_ERR(handle)) {
715 ret = PTR_ERR(handle);
721 /* must not update i_size! */
722 ret = block_commit_write(page, offset, offset);
729 ocfs2_commit_trans(OCFS2_SB(inode->i_sb), handle);
732 page_cache_release(page);
737 static int ocfs2_zero_extend(struct inode *inode,
742 struct super_block *sb = inode->i_sb;
744 start_off = ocfs2_align_bytes_to_blocks(sb, i_size_read(inode));
745 while (start_off < zero_to_size) {
746 ret = ocfs2_write_zero_page(inode, start_off);
752 start_off += sb->s_blocksize;
755 * Very large extends have the potential to lock up
756 * the cpu for extended periods of time.
766 * A tail_to_skip value > 0 indicates that we're being called from
767 * ocfs2_file_aio_write(). This has the following implications:
769 * - we don't want to update i_size
770 * - di_bh will be NULL, which is fine because it's only used in the
771 * case where we want to update i_size.
772 * - ocfs2_zero_extend() will then only be filling the hole created
773 * between i_size and the start of the write.
775 static int ocfs2_extend_file(struct inode *inode,
776 struct buffer_head *di_bh,
783 BUG_ON(!tail_to_skip && !di_bh);
785 /* setattr sometimes calls us like this. */
789 if (i_size_read(inode) == new_i_size)
791 BUG_ON(new_i_size < i_size_read(inode));
793 clusters_to_add = ocfs2_clusters_for_bytes(inode->i_sb, new_i_size) -
794 OCFS2_I(inode)->ip_clusters;
797 * protect the pages that ocfs2_zero_extend is going to be
798 * pulling into the page cache.. we do this before the
799 * metadata extend so that we don't get into the situation
800 * where we've extended the metadata but can't get the data
803 ret = ocfs2_data_lock(inode, 1);
809 if (clusters_to_add) {
810 ret = ocfs2_extend_allocation(inode, clusters_to_add);
818 * Call this even if we don't add any clusters to the tree. We
819 * still need to zero the area between the old i_size and the
822 ret = ocfs2_zero_extend(inode, (u64)new_i_size - tail_to_skip);
829 /* We're being called from ocfs2_setattr() which wants
830 * us to update i_size */
831 ret = ocfs2_simple_size_update(inode, di_bh, new_i_size);
837 ocfs2_data_unlock(inode, 1);
843 int ocfs2_setattr(struct dentry *dentry, struct iattr *attr)
845 int status = 0, size_change;
846 struct inode *inode = dentry->d_inode;
847 struct super_block *sb = inode->i_sb;
848 struct ocfs2_super *osb = OCFS2_SB(sb);
849 struct buffer_head *bh = NULL;
850 handle_t *handle = NULL;
852 mlog_entry("(0x%p, '%.*s')\n", dentry,
853 dentry->d_name.len, dentry->d_name.name);
855 if (attr->ia_valid & ATTR_MODE)
856 mlog(0, "mode change: %d\n", attr->ia_mode);
857 if (attr->ia_valid & ATTR_UID)
858 mlog(0, "uid change: %d\n", attr->ia_uid);
859 if (attr->ia_valid & ATTR_GID)
860 mlog(0, "gid change: %d\n", attr->ia_gid);
861 if (attr->ia_valid & ATTR_SIZE)
862 mlog(0, "size change...\n");
863 if (attr->ia_valid & (ATTR_ATIME | ATTR_MTIME | ATTR_CTIME))
864 mlog(0, "time change...\n");
866 #define OCFS2_VALID_ATTRS (ATTR_ATIME | ATTR_MTIME | ATTR_CTIME | ATTR_SIZE \
867 | ATTR_GID | ATTR_UID | ATTR_MODE)
868 if (!(attr->ia_valid & OCFS2_VALID_ATTRS)) {
869 mlog(0, "can't handle attrs: 0x%x\n", attr->ia_valid);
873 status = inode_change_ok(inode, attr);
877 size_change = S_ISREG(inode->i_mode) && attr->ia_valid & ATTR_SIZE;
879 status = ocfs2_rw_lock(inode, 1);
886 status = ocfs2_meta_lock(inode, &bh, 1);
888 if (status != -ENOENT)
893 if (size_change && attr->ia_size != i_size_read(inode)) {
894 if (i_size_read(inode) > attr->ia_size)
895 status = ocfs2_truncate_file(inode, bh, attr->ia_size);
897 status = ocfs2_extend_file(inode, bh, attr->ia_size, 0);
899 if (status != -ENOSPC)
906 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
907 if (IS_ERR(handle)) {
908 status = PTR_ERR(handle);
913 status = inode_setattr(inode, attr);
919 status = ocfs2_mark_inode_dirty(handle, inode, bh);
924 ocfs2_commit_trans(osb, handle);
926 ocfs2_meta_unlock(inode, 1);
929 ocfs2_rw_unlock(inode, 1);
938 int ocfs2_getattr(struct vfsmount *mnt,
939 struct dentry *dentry,
942 struct inode *inode = dentry->d_inode;
943 struct super_block *sb = dentry->d_inode->i_sb;
944 struct ocfs2_super *osb = sb->s_fs_info;
949 err = ocfs2_inode_revalidate(dentry);
956 generic_fillattr(inode, stat);
958 /* We set the blksize from the cluster size for performance */
959 stat->blksize = osb->s_clustersize;
967 int ocfs2_permission(struct inode *inode, int mask, struct nameidata *nd)
973 ret = ocfs2_meta_lock(inode, NULL, 0);
979 ret = generic_permission(inode, mask, NULL);
981 ocfs2_meta_unlock(inode, 0);
987 static int ocfs2_write_remove_suid(struct inode *inode)
990 struct buffer_head *bh = NULL;
991 struct ocfs2_inode_info *oi = OCFS2_I(inode);
993 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
994 struct ocfs2_dinode *di;
996 mlog_entry("(Inode %llu, mode 0%o)\n",
997 (unsigned long long)oi->ip_blkno, inode->i_mode);
999 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1000 if (handle == NULL) {
1006 ret = ocfs2_read_block(osb, oi->ip_blkno, &bh, OCFS2_BH_CACHED, inode);
1012 ret = ocfs2_journal_access(handle, inode, bh,
1013 OCFS2_JOURNAL_ACCESS_WRITE);
1019 inode->i_mode &= ~S_ISUID;
1020 if ((inode->i_mode & S_ISGID) && (inode->i_mode & S_IXGRP))
1021 inode->i_mode &= ~S_ISGID;
1023 di = (struct ocfs2_dinode *) bh->b_data;
1024 di->i_mode = cpu_to_le16(inode->i_mode);
1026 ret = ocfs2_journal_dirty(handle, bh);
1032 ocfs2_commit_trans(osb, handle);
1038 static int ocfs2_prepare_inode_for_write(struct dentry *dentry,
1043 int ret = 0, meta_level = appending;
1044 struct inode *inode = dentry->d_inode;
1046 loff_t newsize, saved_pos;
1049 * We sample i_size under a read level meta lock to see if our write
1050 * is extending the file, if it is we back off and get a write level
1054 ret = ocfs2_meta_lock(inode, NULL, meta_level);
1061 /* Clear suid / sgid if necessary. We do this here
1062 * instead of later in the write path because
1063 * remove_suid() calls ->setattr without any hint that
1064 * we may have already done our cluster locking. Since
1065 * ocfs2_setattr() *must* take cluster locks to
1066 * proceeed, this will lead us to recursively lock the
1067 * inode. There's also the dinode i_size state which
1068 * can be lost via setattr during extending writes (we
1069 * set inode->i_size at the end of a write. */
1070 if (should_remove_suid(dentry)) {
1071 if (meta_level == 0) {
1072 ocfs2_meta_unlock(inode, meta_level);
1077 ret = ocfs2_write_remove_suid(inode);
1084 /* work on a copy of ppos until we're sure that we won't have
1085 * to recalculate it due to relocking. */
1087 saved_pos = i_size_read(inode);
1088 mlog(0, "O_APPEND: inode->i_size=%llu\n", saved_pos);
1092 newsize = count + saved_pos;
1094 mlog(0, "pos=%lld newsize=%lld cursize=%lld\n",
1095 (long long) saved_pos, (long long) newsize,
1096 (long long) i_size_read(inode));
1098 /* No need for a higher level metadata lock if we're
1099 * never going past i_size. */
1100 if (newsize <= i_size_read(inode))
1103 if (meta_level == 0) {
1104 ocfs2_meta_unlock(inode, meta_level);
1109 spin_lock(&OCFS2_I(inode)->ip_lock);
1110 clusters = ocfs2_clusters_for_bytes(inode->i_sb, newsize) -
1111 OCFS2_I(inode)->ip_clusters;
1112 spin_unlock(&OCFS2_I(inode)->ip_lock);
1114 mlog(0, "Writing at EOF, may need more allocation: "
1115 "i_size = %lld, newsize = %lld, need %u clusters\n",
1116 (long long) i_size_read(inode), (long long) newsize,
1119 /* We only want to continue the rest of this loop if
1120 * our extend will actually require more
1125 ret = ocfs2_extend_file(inode, NULL, newsize, count);
1138 ocfs2_meta_unlock(inode, meta_level);
1144 static ssize_t ocfs2_file_aio_write(struct kiocb *iocb,
1145 const struct iovec *iov,
1146 unsigned long nr_segs,
1149 int ret, rw_level, have_alloc_sem = 0;
1150 struct file *filp = iocb->ki_filp;
1151 struct inode *inode = filp->f_path.dentry->d_inode;
1152 int appending = filp->f_flags & O_APPEND ? 1 : 0;
1154 mlog_entry("(0x%p, %u, '%.*s')\n", filp,
1155 (unsigned int)nr_segs,
1156 filp->f_path.dentry->d_name.len,
1157 filp->f_path.dentry->d_name.name);
1159 /* happy write of zero bytes */
1160 if (iocb->ki_left == 0)
1163 mutex_lock(&inode->i_mutex);
1164 /* to match setattr's i_mutex -> i_alloc_sem -> rw_lock ordering */
1165 if (filp->f_flags & O_DIRECT) {
1167 down_read(&inode->i_alloc_sem);
1170 /* concurrent O_DIRECT writes are allowed */
1171 rw_level = (filp->f_flags & O_DIRECT) ? 0 : 1;
1172 ret = ocfs2_rw_lock(inode, rw_level);
1179 ret = ocfs2_prepare_inode_for_write(filp->f_path.dentry, &iocb->ki_pos,
1180 iocb->ki_left, appending);
1186 /* communicate with ocfs2_dio_end_io */
1187 ocfs2_iocb_set_rw_locked(iocb);
1189 ret = generic_file_aio_write_nolock(iocb, iov, nr_segs, iocb->ki_pos);
1191 /* buffered aio wouldn't have proper lock coverage today */
1192 BUG_ON(ret == -EIOCBQUEUED && !(filp->f_flags & O_DIRECT));
1195 * deep in g_f_a_w_n()->ocfs2_direct_IO we pass in a ocfs2_dio_end_io
1196 * function pointer which is called when o_direct io completes so that
1197 * it can unlock our rw lock. (it's the clustered equivalent of
1198 * i_alloc_sem; protects truncate from racing with pending ios).
1199 * Unfortunately there are error cases which call end_io and others
1200 * that don't. so we don't have to unlock the rw_lock if either an
1201 * async dio is going to do it in the future or an end_io after an
1202 * error has already done it.
1204 if (ret == -EIOCBQUEUED || !ocfs2_iocb_is_rw_locked(iocb)) {
1211 up_read(&inode->i_alloc_sem);
1213 ocfs2_rw_unlock(inode, rw_level);
1214 mutex_unlock(&inode->i_mutex);
1220 static ssize_t ocfs2_file_splice_write(struct pipe_inode_info *pipe,
1227 struct inode *inode = out->f_path.dentry->d_inode;
1229 mlog_entry("(0x%p, 0x%p, %u, '%.*s')\n", out, pipe,
1231 out->f_path.dentry->d_name.len,
1232 out->f_path.dentry->d_name.name);
1234 inode_double_lock(inode, pipe->inode);
1236 ret = ocfs2_rw_lock(inode, 1);
1242 ret = ocfs2_prepare_inode_for_write(out->f_path.dentry, ppos, len, 0);
1248 /* ok, we're done with i_size and alloc work */
1249 ret = generic_file_splice_write_nolock(pipe, out, ppos, len, flags);
1252 ocfs2_rw_unlock(inode, 1);
1254 inode_double_unlock(inode, pipe->inode);
1260 static ssize_t ocfs2_file_splice_read(struct file *in,
1262 struct pipe_inode_info *pipe,
1267 struct inode *inode = in->f_path.dentry->d_inode;
1269 mlog_entry("(0x%p, 0x%p, %u, '%.*s')\n", in, pipe,
1271 in->f_path.dentry->d_name.len,
1272 in->f_path.dentry->d_name.name);
1275 * See the comment in ocfs2_file_aio_read()
1277 ret = ocfs2_meta_lock(inode, NULL, 0);
1282 ocfs2_meta_unlock(inode, 0);
1284 ret = generic_file_splice_read(in, ppos, pipe, len, flags);
1291 static ssize_t ocfs2_file_aio_read(struct kiocb *iocb,
1292 const struct iovec *iov,
1293 unsigned long nr_segs,
1296 int ret = 0, rw_level = -1, have_alloc_sem = 0, lock_level = 0;
1297 struct file *filp = iocb->ki_filp;
1298 struct inode *inode = filp->f_path.dentry->d_inode;
1300 mlog_entry("(0x%p, %u, '%.*s')\n", filp,
1301 (unsigned int)nr_segs,
1302 filp->f_path.dentry->d_name.len,
1303 filp->f_path.dentry->d_name.name);
1312 * buffered reads protect themselves in ->readpage(). O_DIRECT reads
1313 * need locks to protect pending reads from racing with truncate.
1315 if (filp->f_flags & O_DIRECT) {
1316 down_read(&inode->i_alloc_sem);
1319 ret = ocfs2_rw_lock(inode, 0);
1325 /* communicate with ocfs2_dio_end_io */
1326 ocfs2_iocb_set_rw_locked(iocb);
1330 * We're fine letting folks race truncates and extending
1331 * writes with read across the cluster, just like they can
1332 * locally. Hence no rw_lock during read.
1334 * Take and drop the meta data lock to update inode fields
1335 * like i_size. This allows the checks down below
1336 * generic_file_aio_read() a chance of actually working.
1338 ret = ocfs2_meta_lock_atime(inode, filp->f_vfsmnt, &lock_level);
1343 ocfs2_meta_unlock(inode, lock_level);
1345 ret = generic_file_aio_read(iocb, iov, nr_segs, iocb->ki_pos);
1347 mlog(ML_ERROR, "generic_file_aio_read returned -EINVAL\n");
1349 /* buffered aio wouldn't have proper lock coverage today */
1350 BUG_ON(ret == -EIOCBQUEUED && !(filp->f_flags & O_DIRECT));
1352 /* see ocfs2_file_aio_write */
1353 if (ret == -EIOCBQUEUED || !ocfs2_iocb_is_rw_locked(iocb)) {
1360 up_read(&inode->i_alloc_sem);
1362 ocfs2_rw_unlock(inode, rw_level);
1368 const struct inode_operations ocfs2_file_iops = {
1369 .setattr = ocfs2_setattr,
1370 .getattr = ocfs2_getattr,
1371 .permission = ocfs2_permission,
1374 const struct inode_operations ocfs2_special_file_iops = {
1375 .setattr = ocfs2_setattr,
1376 .getattr = ocfs2_getattr,
1377 .permission = ocfs2_permission,
1380 const struct file_operations ocfs2_fops = {
1381 .read = do_sync_read,
1382 .write = do_sync_write,
1383 .sendfile = generic_file_sendfile,
1385 .fsync = ocfs2_sync_file,
1386 .release = ocfs2_file_release,
1387 .open = ocfs2_file_open,
1388 .aio_read = ocfs2_file_aio_read,
1389 .aio_write = ocfs2_file_aio_write,
1390 .ioctl = ocfs2_ioctl,
1391 .splice_read = ocfs2_file_splice_read,
1392 .splice_write = ocfs2_file_splice_write,
1395 const struct file_operations ocfs2_dops = {
1396 .read = generic_read_dir,
1397 .readdir = ocfs2_readdir,
1398 .fsync = ocfs2_sync_file,
1399 .ioctl = ocfs2_ioctl,