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>
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 static 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_inode_sector_count(inode);
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,
265 struct ocfs2_dinode *di;
269 /* TODO: This needs to actually orphan the inode in this
272 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
273 if (IS_ERR(handle)) {
274 status = PTR_ERR(handle);
279 status = ocfs2_journal_access(handle, inode, fe_bh,
280 OCFS2_JOURNAL_ACCESS_WRITE);
287 * Do this before setting i_size.
289 status = ocfs2_zero_tail_for_truncate(inode, handle, new_i_size);
295 i_size_write(inode, new_i_size);
296 inode->i_blocks = ocfs2_align_bytes_to_sectors(new_i_size);
297 inode->i_ctime = inode->i_mtime = CURRENT_TIME;
299 di = (struct ocfs2_dinode *) fe_bh->b_data;
300 di->i_size = cpu_to_le64(new_i_size);
301 di->i_ctime = di->i_mtime = cpu_to_le64(inode->i_ctime.tv_sec);
302 di->i_ctime_nsec = di->i_mtime_nsec = cpu_to_le32(inode->i_ctime.tv_nsec);
304 status = ocfs2_journal_dirty(handle, fe_bh);
309 ocfs2_commit_trans(osb, handle);
316 static int ocfs2_truncate_file(struct inode *inode,
317 struct buffer_head *di_bh,
321 struct ocfs2_dinode *fe = NULL;
322 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
323 struct ocfs2_truncate_context *tc = NULL;
325 mlog_entry("(inode = %llu, new_i_size = %llu\n",
326 (unsigned long long)OCFS2_I(inode)->ip_blkno,
327 (unsigned long long)new_i_size);
329 fe = (struct ocfs2_dinode *) di_bh->b_data;
330 if (!OCFS2_IS_VALID_DINODE(fe)) {
331 OCFS2_RO_ON_INVALID_DINODE(inode->i_sb, fe);
336 mlog_bug_on_msg(le64_to_cpu(fe->i_size) != i_size_read(inode),
337 "Inode %llu, inode i_size = %lld != di "
338 "i_size = %llu, i_flags = 0x%x\n",
339 (unsigned long long)OCFS2_I(inode)->ip_blkno,
341 (unsigned long long)le64_to_cpu(fe->i_size),
342 le32_to_cpu(fe->i_flags));
344 if (new_i_size > le64_to_cpu(fe->i_size)) {
345 mlog(0, "asked to truncate file with size (%llu) to size (%llu)!\n",
346 (unsigned long long)le64_to_cpu(fe->i_size),
347 (unsigned long long)new_i_size);
353 mlog(0, "inode %llu, i_size = %llu, new_i_size = %llu\n",
354 (unsigned long long)le64_to_cpu(fe->i_blkno),
355 (unsigned long long)le64_to_cpu(fe->i_size),
356 (unsigned long long)new_i_size);
358 /* lets handle the simple truncate cases before doing any more
359 * cluster locking. */
360 if (new_i_size == le64_to_cpu(fe->i_size))
363 down_write(&OCFS2_I(inode)->ip_alloc_sem);
365 /* This forces other nodes to sync and drop their pages. Do
366 * this even if we have a truncate without allocation change -
367 * ocfs2 cluster sizes can be much greater than page size, so
368 * we have to truncate them anyway. */
369 status = ocfs2_data_lock(inode, 1);
371 up_write(&OCFS2_I(inode)->ip_alloc_sem);
377 unmap_mapping_range(inode->i_mapping, new_i_size + PAGE_SIZE - 1, 0, 1);
378 truncate_inode_pages(inode->i_mapping, new_i_size);
380 /* alright, we're going to need to do a full blown alloc size
381 * change. Orphan the inode so that recovery can complete the
382 * truncate if necessary. This does the task of marking
384 status = ocfs2_orphan_for_truncate(osb, inode, di_bh, new_i_size);
387 goto bail_unlock_data;
390 status = ocfs2_prepare_truncate(osb, inode, di_bh, &tc);
393 goto bail_unlock_data;
396 status = ocfs2_commit_truncate(osb, inode, di_bh, tc);
399 goto bail_unlock_data;
402 /* TODO: orphan dir cleanup here. */
404 ocfs2_data_unlock(inode, 1);
406 up_write(&OCFS2_I(inode)->ip_alloc_sem);
415 * extend allocation only here.
416 * we'll update all the disk stuff, and oip->alloc_size
418 * expect stuff to be locked, a transaction started and enough data /
419 * metadata reservations in the contexts.
421 * Will return -EAGAIN, and a reason if a restart is needed.
422 * If passed in, *reason will always be set, even in error.
424 int ocfs2_do_extend_allocation(struct ocfs2_super *osb,
428 struct buffer_head *fe_bh,
430 struct ocfs2_alloc_context *data_ac,
431 struct ocfs2_alloc_context *meta_ac,
432 enum ocfs2_alloc_restarted *reason_ret)
436 struct ocfs2_dinode *fe = (struct ocfs2_dinode *) fe_bh->b_data;
437 enum ocfs2_alloc_restarted reason = RESTART_NONE;
438 u32 bit_off, num_bits;
441 BUG_ON(!clusters_to_add);
443 free_extents = ocfs2_num_free_extents(osb, inode, fe);
444 if (free_extents < 0) {
445 status = free_extents;
450 /* there are two cases which could cause us to EAGAIN in the
451 * we-need-more-metadata case:
452 * 1) we haven't reserved *any*
453 * 2) we are so fragmented, we've needed to add metadata too
455 if (!free_extents && !meta_ac) {
456 mlog(0, "we haven't reserved any metadata!\n");
458 reason = RESTART_META;
460 } else if ((!free_extents)
461 && (ocfs2_alloc_context_bits_left(meta_ac)
462 < ocfs2_extend_meta_needed(fe))) {
463 mlog(0, "filesystem is really fragmented...\n");
465 reason = RESTART_META;
469 status = ocfs2_claim_clusters(osb, handle, data_ac, 1,
470 &bit_off, &num_bits);
472 if (status != -ENOSPC)
477 BUG_ON(num_bits > clusters_to_add);
479 /* reserve our write early -- insert_extent may update the inode */
480 status = ocfs2_journal_access(handle, inode, fe_bh,
481 OCFS2_JOURNAL_ACCESS_WRITE);
487 block = ocfs2_clusters_to_blocks(osb->sb, bit_off);
488 mlog(0, "Allocating %u clusters at block %u for inode %llu\n",
489 num_bits, bit_off, (unsigned long long)OCFS2_I(inode)->ip_blkno);
490 status = ocfs2_insert_extent(osb, handle, inode, fe_bh,
491 *logical_offset, block, num_bits,
498 status = ocfs2_journal_dirty(handle, fe_bh);
504 clusters_to_add -= num_bits;
505 *logical_offset += num_bits;
507 if (clusters_to_add) {
508 mlog(0, "need to alloc once more, clusters = %u, wanted = "
509 "%u\n", fe->i_clusters, clusters_to_add);
511 reason = RESTART_TRANS;
517 *reason_ret = reason;
522 * For a given allocation, determine which allocators will need to be
523 * accessed, and lock them, reserving the appropriate number of bits.
525 * Called from ocfs2_extend_allocation() for file systems which don't
526 * support holes, and from ocfs2_write() for file systems which
527 * understand sparse inodes.
529 int ocfs2_lock_allocators(struct inode *inode, struct ocfs2_dinode *di,
531 struct ocfs2_alloc_context **data_ac,
532 struct ocfs2_alloc_context **meta_ac)
534 int ret, num_free_extents;
535 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
540 mlog(0, "extend inode %llu, i_size = %lld, di->i_clusters = %u, "
541 "clusters_to_add = %u\n",
542 (unsigned long long)OCFS2_I(inode)->ip_blkno, i_size_read(inode),
543 le32_to_cpu(di->i_clusters), clusters_to_add);
545 num_free_extents = ocfs2_num_free_extents(osb, inode, di);
546 if (num_free_extents < 0) {
547 ret = num_free_extents;
553 * Sparse allocation file systems need to be more conservative
554 * with reserving room for expansion - the actual allocation
555 * happens while we've got a journal handle open so re-taking
556 * a cluster lock (because we ran out of room for another
557 * extent) will violate ordering rules.
559 * Most of the time we'll only be seeing this 1 cluster at a time
562 if (!num_free_extents ||
563 (ocfs2_sparse_alloc(osb) && num_free_extents < clusters_to_add)) {
564 ret = ocfs2_reserve_new_metadata(osb, di, meta_ac);
572 ret = ocfs2_reserve_clusters(osb, clusters_to_add, data_ac);
582 ocfs2_free_alloc_context(*meta_ac);
587 * We cannot have an error and a non null *data_ac.
594 static int ocfs2_extend_allocation(struct inode *inode,
598 int restart_func = 0;
599 int drop_alloc_sem = 0;
601 u32 prev_clusters, logical_start;
602 struct buffer_head *bh = NULL;
603 struct ocfs2_dinode *fe = NULL;
604 handle_t *handle = NULL;
605 struct ocfs2_alloc_context *data_ac = NULL;
606 struct ocfs2_alloc_context *meta_ac = NULL;
607 enum ocfs2_alloc_restarted why;
608 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
610 mlog_entry("(clusters_to_add = %u)\n", clusters_to_add);
613 * This function only exists for file systems which don't
616 BUG_ON(ocfs2_sparse_alloc(osb));
618 status = ocfs2_read_block(osb, OCFS2_I(inode)->ip_blkno, &bh,
619 OCFS2_BH_CACHED, inode);
625 fe = (struct ocfs2_dinode *) bh->b_data;
626 if (!OCFS2_IS_VALID_DINODE(fe)) {
627 OCFS2_RO_ON_INVALID_DINODE(inode->i_sb, fe);
632 logical_start = OCFS2_I(inode)->ip_clusters;
635 BUG_ON(le32_to_cpu(fe->i_clusters) != OCFS2_I(inode)->ip_clusters);
637 /* blocks peope in read/write from reading our allocation
638 * until we're done changing it. We depend on i_mutex to block
639 * other extend/truncate calls while we're here. Ordering wrt
640 * start_trans is important here -- always do it before! */
641 down_write(&OCFS2_I(inode)->ip_alloc_sem);
644 status = ocfs2_lock_allocators(inode, fe, clusters_to_add, &data_ac,
651 credits = ocfs2_calc_extend_credits(osb->sb, fe, clusters_to_add);
652 handle = ocfs2_start_trans(osb, credits);
653 if (IS_ERR(handle)) {
654 status = PTR_ERR(handle);
660 restarted_transaction:
661 /* reserve a write to the file entry early on - that we if we
662 * run out of credits in the allocation path, we can still
664 status = ocfs2_journal_access(handle, inode, bh,
665 OCFS2_JOURNAL_ACCESS_WRITE);
671 prev_clusters = OCFS2_I(inode)->ip_clusters;
673 status = ocfs2_do_extend_allocation(osb,
682 if ((status < 0) && (status != -EAGAIN)) {
683 if (status != -ENOSPC)
688 status = ocfs2_journal_dirty(handle, bh);
694 spin_lock(&OCFS2_I(inode)->ip_lock);
695 clusters_to_add -= (OCFS2_I(inode)->ip_clusters - prev_clusters);
696 spin_unlock(&OCFS2_I(inode)->ip_lock);
698 if (why != RESTART_NONE && clusters_to_add) {
699 if (why == RESTART_META) {
700 mlog(0, "restarting function.\n");
703 BUG_ON(why != RESTART_TRANS);
705 mlog(0, "restarting transaction.\n");
706 /* TODO: This can be more intelligent. */
707 credits = ocfs2_calc_extend_credits(osb->sb,
710 status = ocfs2_extend_trans(handle, credits);
712 /* handle still has to be committed at
718 goto restarted_transaction;
722 mlog(0, "fe: i_clusters = %u, i_size=%llu\n",
723 le32_to_cpu(fe->i_clusters),
724 (unsigned long long)le64_to_cpu(fe->i_size));
725 mlog(0, "inode: ip_clusters=%u, i_size=%lld\n",
726 OCFS2_I(inode)->ip_clusters, i_size_read(inode));
729 if (drop_alloc_sem) {
730 up_write(&OCFS2_I(inode)->ip_alloc_sem);
734 ocfs2_commit_trans(osb, handle);
738 ocfs2_free_alloc_context(data_ac);
742 ocfs2_free_alloc_context(meta_ac);
745 if ((!status) && restart_func) {
758 /* Some parts of this taken from generic_cont_expand, which turned out
759 * to be too fragile to do exactly what we need without us having to
760 * worry about recursive locking in ->prepare_write() and
761 * ->commit_write(). */
762 static int ocfs2_write_zero_page(struct inode *inode,
765 struct address_space *mapping = inode->i_mapping;
769 handle_t *handle = NULL;
772 offset = (size & (PAGE_CACHE_SIZE-1)); /* Within page */
773 /* ugh. in prepare/commit_write, if from==to==start of block, we
774 ** skip the prepare. make sure we never send an offset for the start
777 if ((offset & (inode->i_sb->s_blocksize - 1)) == 0) {
780 index = size >> PAGE_CACHE_SHIFT;
782 page = grab_cache_page(mapping, index);
789 ret = ocfs2_prepare_write_nolock(inode, page, offset, offset);
795 if (ocfs2_should_order_data(inode)) {
796 handle = ocfs2_start_walk_page_trans(inode, page, offset,
798 if (IS_ERR(handle)) {
799 ret = PTR_ERR(handle);
805 /* must not update i_size! */
806 ret = block_commit_write(page, offset, offset);
813 ocfs2_commit_trans(OCFS2_SB(inode->i_sb), handle);
816 page_cache_release(page);
821 static int ocfs2_zero_extend(struct inode *inode,
826 struct super_block *sb = inode->i_sb;
828 start_off = ocfs2_align_bytes_to_blocks(sb, i_size_read(inode));
829 while (start_off < zero_to_size) {
830 ret = ocfs2_write_zero_page(inode, start_off);
836 start_off += sb->s_blocksize;
839 * Very large extends have the potential to lock up
840 * the cpu for extended periods of time.
850 * A tail_to_skip value > 0 indicates that we're being called from
851 * ocfs2_file_aio_write(). This has the following implications:
853 * - we don't want to update i_size
854 * - di_bh will be NULL, which is fine because it's only used in the
855 * case where we want to update i_size.
856 * - ocfs2_zero_extend() will then only be filling the hole created
857 * between i_size and the start of the write.
859 static int ocfs2_extend_file(struct inode *inode,
860 struct buffer_head *di_bh,
865 u32 clusters_to_add = 0;
867 BUG_ON(!tail_to_skip && !di_bh);
869 /* setattr sometimes calls us like this. */
873 if (i_size_read(inode) == new_i_size)
875 BUG_ON(new_i_size < i_size_read(inode));
877 if (ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb))) {
878 BUG_ON(tail_to_skip != 0);
879 goto out_update_size;
882 clusters_to_add = ocfs2_clusters_for_bytes(inode->i_sb, new_i_size) -
883 OCFS2_I(inode)->ip_clusters;
886 * protect the pages that ocfs2_zero_extend is going to be
887 * pulling into the page cache.. we do this before the
888 * metadata extend so that we don't get into the situation
889 * where we've extended the metadata but can't get the data
892 ret = ocfs2_data_lock(inode, 1);
898 if (clusters_to_add) {
899 ret = ocfs2_extend_allocation(inode, clusters_to_add);
907 * Call this even if we don't add any clusters to the tree. We
908 * still need to zero the area between the old i_size and the
911 ret = ocfs2_zero_extend(inode, (u64)new_i_size - tail_to_skip);
919 /* We're being called from ocfs2_setattr() which wants
920 * us to update i_size */
921 ret = ocfs2_simple_size_update(inode, di_bh, new_i_size);
927 if (!ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)))
928 ocfs2_data_unlock(inode, 1);
934 int ocfs2_setattr(struct dentry *dentry, struct iattr *attr)
936 int status = 0, size_change;
937 struct inode *inode = dentry->d_inode;
938 struct super_block *sb = inode->i_sb;
939 struct ocfs2_super *osb = OCFS2_SB(sb);
940 struct buffer_head *bh = NULL;
941 handle_t *handle = NULL;
943 mlog_entry("(0x%p, '%.*s')\n", dentry,
944 dentry->d_name.len, dentry->d_name.name);
946 if (attr->ia_valid & ATTR_MODE)
947 mlog(0, "mode change: %d\n", attr->ia_mode);
948 if (attr->ia_valid & ATTR_UID)
949 mlog(0, "uid change: %d\n", attr->ia_uid);
950 if (attr->ia_valid & ATTR_GID)
951 mlog(0, "gid change: %d\n", attr->ia_gid);
952 if (attr->ia_valid & ATTR_SIZE)
953 mlog(0, "size change...\n");
954 if (attr->ia_valid & (ATTR_ATIME | ATTR_MTIME | ATTR_CTIME))
955 mlog(0, "time change...\n");
957 #define OCFS2_VALID_ATTRS (ATTR_ATIME | ATTR_MTIME | ATTR_CTIME | ATTR_SIZE \
958 | ATTR_GID | ATTR_UID | ATTR_MODE)
959 if (!(attr->ia_valid & OCFS2_VALID_ATTRS)) {
960 mlog(0, "can't handle attrs: 0x%x\n", attr->ia_valid);
964 status = inode_change_ok(inode, attr);
968 size_change = S_ISREG(inode->i_mode) && attr->ia_valid & ATTR_SIZE;
970 status = ocfs2_rw_lock(inode, 1);
977 status = ocfs2_meta_lock(inode, &bh, 1);
979 if (status != -ENOENT)
984 if (size_change && attr->ia_size != i_size_read(inode)) {
985 if (i_size_read(inode) > attr->ia_size)
986 status = ocfs2_truncate_file(inode, bh, attr->ia_size);
988 status = ocfs2_extend_file(inode, bh, attr->ia_size, 0);
990 if (status != -ENOSPC)
997 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
998 if (IS_ERR(handle)) {
999 status = PTR_ERR(handle);
1005 * This will intentionally not wind up calling vmtruncate(),
1006 * since all the work for a size change has been done above.
1007 * Otherwise, we could get into problems with truncate as
1008 * ip_alloc_sem is used there to protect against i_size
1011 status = inode_setattr(inode, attr);
1017 status = ocfs2_mark_inode_dirty(handle, inode, bh);
1022 ocfs2_commit_trans(osb, handle);
1024 ocfs2_meta_unlock(inode, 1);
1027 ocfs2_rw_unlock(inode, 1);
1036 int ocfs2_getattr(struct vfsmount *mnt,
1037 struct dentry *dentry,
1040 struct inode *inode = dentry->d_inode;
1041 struct super_block *sb = dentry->d_inode->i_sb;
1042 struct ocfs2_super *osb = sb->s_fs_info;
1047 err = ocfs2_inode_revalidate(dentry);
1054 generic_fillattr(inode, stat);
1056 /* We set the blksize from the cluster size for performance */
1057 stat->blksize = osb->s_clustersize;
1065 int ocfs2_permission(struct inode *inode, int mask, struct nameidata *nd)
1071 ret = ocfs2_meta_lock(inode, NULL, 0);
1078 ret = generic_permission(inode, mask, NULL);
1080 ocfs2_meta_unlock(inode, 0);
1086 static int ocfs2_write_remove_suid(struct inode *inode)
1089 struct buffer_head *bh = NULL;
1090 struct ocfs2_inode_info *oi = OCFS2_I(inode);
1092 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1093 struct ocfs2_dinode *di;
1095 mlog_entry("(Inode %llu, mode 0%o)\n",
1096 (unsigned long long)oi->ip_blkno, inode->i_mode);
1098 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1099 if (handle == NULL) {
1105 ret = ocfs2_read_block(osb, oi->ip_blkno, &bh, OCFS2_BH_CACHED, inode);
1111 ret = ocfs2_journal_access(handle, inode, bh,
1112 OCFS2_JOURNAL_ACCESS_WRITE);
1118 inode->i_mode &= ~S_ISUID;
1119 if ((inode->i_mode & S_ISGID) && (inode->i_mode & S_IXGRP))
1120 inode->i_mode &= ~S_ISGID;
1122 di = (struct ocfs2_dinode *) bh->b_data;
1123 di->i_mode = cpu_to_le16(inode->i_mode);
1125 ret = ocfs2_journal_dirty(handle, bh);
1131 ocfs2_commit_trans(osb, handle);
1138 * Will look for holes and unwritten extents in the range starting at
1139 * pos for count bytes (inclusive).
1141 static int ocfs2_check_range_for_holes(struct inode *inode, loff_t pos,
1145 unsigned int extent_flags;
1146 u32 cpos, clusters, extent_len, phys_cpos;
1147 struct super_block *sb = inode->i_sb;
1149 cpos = pos >> OCFS2_SB(sb)->s_clustersize_bits;
1150 clusters = ocfs2_clusters_for_bytes(sb, pos + count) - cpos;
1153 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos, &extent_len,
1160 if (phys_cpos == 0 || (extent_flags & OCFS2_EXT_UNWRITTEN)) {
1165 if (extent_len > clusters)
1166 extent_len = clusters;
1168 clusters -= extent_len;
1175 static int ocfs2_prepare_inode_for_write(struct dentry *dentry,
1181 int ret = 0, meta_level = appending;
1182 struct inode *inode = dentry->d_inode;
1184 loff_t newsize, saved_pos;
1187 * We sample i_size under a read level meta lock to see if our write
1188 * is extending the file, if it is we back off and get a write level
1192 ret = ocfs2_meta_lock(inode, NULL, meta_level);
1199 /* Clear suid / sgid if necessary. We do this here
1200 * instead of later in the write path because
1201 * remove_suid() calls ->setattr without any hint that
1202 * we may have already done our cluster locking. Since
1203 * ocfs2_setattr() *must* take cluster locks to
1204 * proceeed, this will lead us to recursively lock the
1205 * inode. There's also the dinode i_size state which
1206 * can be lost via setattr during extending writes (we
1207 * set inode->i_size at the end of a write. */
1208 if (should_remove_suid(dentry)) {
1209 if (meta_level == 0) {
1210 ocfs2_meta_unlock(inode, meta_level);
1215 ret = ocfs2_write_remove_suid(inode);
1222 /* work on a copy of ppos until we're sure that we won't have
1223 * to recalculate it due to relocking. */
1225 saved_pos = i_size_read(inode);
1226 mlog(0, "O_APPEND: inode->i_size=%llu\n", saved_pos);
1231 if (ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb))) {
1232 loff_t end = saved_pos + count;
1235 * Skip the O_DIRECT checks if we don't need
1238 if (!direct_io || !(*direct_io))
1242 * Allowing concurrent direct writes means
1243 * i_size changes wouldn't be synchronized, so
1244 * one node could wind up truncating another
1247 if (end > i_size_read(inode)) {
1253 * We don't fill holes during direct io, so
1254 * check for them here. If any are found, the
1255 * caller will have to retake some cluster
1256 * locks and initiate the io as buffered.
1258 ret = ocfs2_check_range_for_holes(inode, saved_pos,
1269 * The rest of this loop is concerned with legacy file
1270 * systems which don't support sparse files.
1273 newsize = count + saved_pos;
1275 mlog(0, "pos=%lld newsize=%lld cursize=%lld\n",
1276 (long long) saved_pos, (long long) newsize,
1277 (long long) i_size_read(inode));
1279 /* No need for a higher level metadata lock if we're
1280 * never going past i_size. */
1281 if (newsize <= i_size_read(inode))
1284 if (meta_level == 0) {
1285 ocfs2_meta_unlock(inode, meta_level);
1290 spin_lock(&OCFS2_I(inode)->ip_lock);
1291 clusters = ocfs2_clusters_for_bytes(inode->i_sb, newsize) -
1292 OCFS2_I(inode)->ip_clusters;
1293 spin_unlock(&OCFS2_I(inode)->ip_lock);
1295 mlog(0, "Writing at EOF, may need more allocation: "
1296 "i_size = %lld, newsize = %lld, need %u clusters\n",
1297 (long long) i_size_read(inode), (long long) newsize,
1300 /* We only want to continue the rest of this loop if
1301 * our extend will actually require more
1306 ret = ocfs2_extend_file(inode, NULL, newsize, count);
1319 ocfs2_meta_unlock(inode, meta_level);
1326 ocfs2_set_next_iovec(const struct iovec **iovp, size_t *basep, size_t bytes)
1328 const struct iovec *iov = *iovp;
1329 size_t base = *basep;
1332 int copy = min(bytes, iov->iov_len - base);
1336 if (iov->iov_len == base) {
1345 static struct page * ocfs2_get_write_source(char **ret_src_buf,
1346 const struct iovec *cur_iov,
1350 char *buf = cur_iov->iov_base + iov_offset;
1351 struct page *src_page = NULL;
1354 off = (unsigned long)(buf) & ~PAGE_CACHE_MASK;
1356 if (!segment_eq(get_fs(), KERNEL_DS)) {
1358 * Pull in the user page. We want to do this outside
1359 * of the meta data locks in order to preserve locking
1360 * order in case of page fault.
1362 ret = get_user_pages(current, current->mm,
1363 (unsigned long)buf & PAGE_CACHE_MASK, 1,
1364 0, 0, &src_page, NULL);
1366 *ret_src_buf = kmap(src_page) + off;
1368 src_page = ERR_PTR(-EFAULT);
1376 static void ocfs2_put_write_source(struct page *page)
1380 page_cache_release(page);
1384 static ssize_t ocfs2_file_buffered_write(struct file *file, loff_t *ppos,
1385 const struct iovec *iov,
1386 unsigned long nr_segs,
1388 ssize_t o_direct_written)
1391 ssize_t copied, total = 0;
1392 size_t iov_offset = 0, bytes;
1394 const struct iovec *cur_iov = iov;
1395 struct page *user_page, *page;
1400 * handle partial DIO write. Adjust cur_iov if needed.
1402 ocfs2_set_next_iovec(&cur_iov, &iov_offset, o_direct_written);
1407 user_page = ocfs2_get_write_source(&buf, cur_iov, iov_offset);
1408 if (IS_ERR(user_page)) {
1409 ret = PTR_ERR(user_page);
1413 /* Stay within our page boundaries */
1414 bytes = min((PAGE_CACHE_SIZE - ((unsigned long)pos & ~PAGE_CACHE_MASK)),
1415 (PAGE_CACHE_SIZE - ((unsigned long)buf & ~PAGE_CACHE_MASK)));
1416 /* Stay within the vector boundary */
1417 bytes = min_t(size_t, bytes, cur_iov->iov_len - iov_offset);
1418 /* Stay within count */
1419 bytes = min(bytes, count);
1422 ret = ocfs2_write_begin(file, file->f_mapping, pos, bytes, 0,
1429 dst = kmap_atomic(page, KM_USER0);
1430 memcpy(dst + (pos & (PAGE_CACHE_SIZE - 1)), buf, bytes);
1431 kunmap_atomic(dst, KM_USER0);
1432 flush_dcache_page(page);
1433 ocfs2_put_write_source(user_page);
1435 copied = ocfs2_write_end(file, file->f_mapping, pos, bytes,
1436 bytes, page, fsdata);
1444 *ppos = pos + copied;
1447 ocfs2_set_next_iovec(&cur_iov, &iov_offset, copied);
1451 return total ? total : ret;
1454 static ssize_t ocfs2_file_aio_write(struct kiocb *iocb,
1455 const struct iovec *iov,
1456 unsigned long nr_segs,
1459 int ret, direct_io, appending, rw_level, have_alloc_sem = 0;
1460 int can_do_direct, sync = 0;
1461 ssize_t written = 0;
1462 size_t ocount; /* original count */
1463 size_t count; /* after file limit checks */
1464 loff_t *ppos = &iocb->ki_pos;
1465 struct file *file = iocb->ki_filp;
1466 struct inode *inode = file->f_path.dentry->d_inode;
1468 mlog_entry("(0x%p, %u, '%.*s')\n", file,
1469 (unsigned int)nr_segs,
1470 file->f_path.dentry->d_name.len,
1471 file->f_path.dentry->d_name.name);
1473 if (iocb->ki_left == 0)
1476 ret = generic_segment_checks(iov, &nr_segs, &ocount, VERIFY_READ);
1482 vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
1484 appending = file->f_flags & O_APPEND ? 1 : 0;
1485 direct_io = file->f_flags & O_DIRECT ? 1 : 0;
1487 mutex_lock(&inode->i_mutex);
1490 /* to match setattr's i_mutex -> i_alloc_sem -> rw_lock ordering */
1492 down_read(&inode->i_alloc_sem);
1496 /* concurrent O_DIRECT writes are allowed */
1497 rw_level = !direct_io;
1498 ret = ocfs2_rw_lock(inode, rw_level);
1504 can_do_direct = direct_io;
1505 ret = ocfs2_prepare_inode_for_write(file->f_path.dentry, ppos,
1506 iocb->ki_left, appending,
1514 * We can't complete the direct I/O as requested, fall back to
1517 if (direct_io && !can_do_direct) {
1518 ocfs2_rw_unlock(inode, rw_level);
1519 up_read(&inode->i_alloc_sem);
1529 if (!sync && ((file->f_flags & O_SYNC) || IS_SYNC(inode)))
1533 * XXX: Is it ok to execute these checks a second time?
1535 ret = generic_write_checks(file, ppos, &count, S_ISBLK(inode->i_mode));
1540 * Set pos so that sync_page_range_nolock() below understands
1541 * where to start from. We might've moved it around via the
1542 * calls above. The range we want to actually sync starts from
1548 /* communicate with ocfs2_dio_end_io */
1549 ocfs2_iocb_set_rw_locked(iocb, rw_level);
1552 written = generic_file_direct_write(iocb, iov, &nr_segs, *ppos,
1553 ppos, count, ocount);
1559 written = ocfs2_file_buffered_write(file, ppos, iov, nr_segs,
1563 if (ret != -EFAULT || ret != -ENOSPC)
1570 /* buffered aio wouldn't have proper lock coverage today */
1571 BUG_ON(ret == -EIOCBQUEUED && !(file->f_flags & O_DIRECT));
1574 * deep in g_f_a_w_n()->ocfs2_direct_IO we pass in a ocfs2_dio_end_io
1575 * function pointer which is called when o_direct io completes so that
1576 * it can unlock our rw lock. (it's the clustered equivalent of
1577 * i_alloc_sem; protects truncate from racing with pending ios).
1578 * Unfortunately there are error cases which call end_io and others
1579 * that don't. so we don't have to unlock the rw_lock if either an
1580 * async dio is going to do it in the future or an end_io after an
1581 * error has already done it.
1583 if (ret == -EIOCBQUEUED || !ocfs2_iocb_is_rw_locked(iocb)) {
1590 ocfs2_rw_unlock(inode, rw_level);
1594 up_read(&inode->i_alloc_sem);
1596 if (written > 0 && sync) {
1599 err = sync_page_range_nolock(inode, file->f_mapping, pos, count);
1604 mutex_unlock(&inode->i_mutex);
1607 return written ? written : ret;
1610 static int ocfs2_splice_write_actor(struct pipe_inode_info *pipe,
1611 struct pipe_buffer *buf,
1612 struct splice_desc *sd)
1616 struct file *file = sd->u.file;
1617 unsigned int offset;
1618 struct page *page = NULL;
1622 ret = buf->ops->confirm(pipe, buf);
1626 offset = sd->pos & ~PAGE_CACHE_MASK;
1628 if (count + offset > PAGE_CACHE_SIZE)
1629 count = PAGE_CACHE_SIZE - offset;
1631 ret = ocfs2_write_begin(file, file->f_mapping, sd->pos, count, 0,
1638 src = buf->ops->map(pipe, buf, 1);
1639 dst = kmap_atomic(page, KM_USER1);
1640 memcpy(dst + offset, src + buf->offset, count);
1641 kunmap_atomic(page, KM_USER1);
1642 buf->ops->unmap(pipe, buf, src);
1644 copied = ocfs2_write_end(file, file->f_mapping, sd->pos, count, count,
1653 return copied ? copied : ret;
1656 static ssize_t __ocfs2_file_splice_write(struct pipe_inode_info *pipe,
1663 struct address_space *mapping = out->f_mapping;
1664 struct inode *inode = mapping->host;
1665 struct splice_desc sd = {
1672 ret = __splice_from_pipe(pipe, &sd, ocfs2_splice_write_actor);
1676 if (unlikely((out->f_flags & O_SYNC) || IS_SYNC(inode))) {
1677 err = generic_osync_inode(inode, mapping,
1678 OSYNC_METADATA|OSYNC_DATA);
1687 static ssize_t ocfs2_file_splice_write(struct pipe_inode_info *pipe,
1694 struct inode *inode = out->f_path.dentry->d_inode;
1696 mlog_entry("(0x%p, 0x%p, %u, '%.*s')\n", out, pipe,
1698 out->f_path.dentry->d_name.len,
1699 out->f_path.dentry->d_name.name);
1701 inode_double_lock(inode, pipe->inode);
1703 ret = ocfs2_rw_lock(inode, 1);
1709 ret = ocfs2_prepare_inode_for_write(out->f_path.dentry, ppos, len, 0,
1716 /* ok, we're done with i_size and alloc work */
1717 ret = __ocfs2_file_splice_write(pipe, out, ppos, len, flags);
1720 ocfs2_rw_unlock(inode, 1);
1722 inode_double_unlock(inode, pipe->inode);
1728 static ssize_t ocfs2_file_splice_read(struct file *in,
1730 struct pipe_inode_info *pipe,
1735 struct inode *inode = in->f_path.dentry->d_inode;
1737 mlog_entry("(0x%p, 0x%p, %u, '%.*s')\n", in, pipe,
1739 in->f_path.dentry->d_name.len,
1740 in->f_path.dentry->d_name.name);
1743 * See the comment in ocfs2_file_aio_read()
1745 ret = ocfs2_meta_lock(inode, NULL, 0);
1750 ocfs2_meta_unlock(inode, 0);
1752 ret = generic_file_splice_read(in, ppos, pipe, len, flags);
1759 static ssize_t ocfs2_file_aio_read(struct kiocb *iocb,
1760 const struct iovec *iov,
1761 unsigned long nr_segs,
1764 int ret = 0, rw_level = -1, have_alloc_sem = 0, lock_level = 0;
1765 struct file *filp = iocb->ki_filp;
1766 struct inode *inode = filp->f_path.dentry->d_inode;
1768 mlog_entry("(0x%p, %u, '%.*s')\n", filp,
1769 (unsigned int)nr_segs,
1770 filp->f_path.dentry->d_name.len,
1771 filp->f_path.dentry->d_name.name);
1780 * buffered reads protect themselves in ->readpage(). O_DIRECT reads
1781 * need locks to protect pending reads from racing with truncate.
1783 if (filp->f_flags & O_DIRECT) {
1784 down_read(&inode->i_alloc_sem);
1787 ret = ocfs2_rw_lock(inode, 0);
1793 /* communicate with ocfs2_dio_end_io */
1794 ocfs2_iocb_set_rw_locked(iocb, rw_level);
1798 * We're fine letting folks race truncates and extending
1799 * writes with read across the cluster, just like they can
1800 * locally. Hence no rw_lock during read.
1802 * Take and drop the meta data lock to update inode fields
1803 * like i_size. This allows the checks down below
1804 * generic_file_aio_read() a chance of actually working.
1806 ret = ocfs2_meta_lock_atime(inode, filp->f_vfsmnt, &lock_level);
1811 ocfs2_meta_unlock(inode, lock_level);
1813 ret = generic_file_aio_read(iocb, iov, nr_segs, iocb->ki_pos);
1815 mlog(ML_ERROR, "generic_file_aio_read returned -EINVAL\n");
1817 /* buffered aio wouldn't have proper lock coverage today */
1818 BUG_ON(ret == -EIOCBQUEUED && !(filp->f_flags & O_DIRECT));
1820 /* see ocfs2_file_aio_write */
1821 if (ret == -EIOCBQUEUED || !ocfs2_iocb_is_rw_locked(iocb)) {
1828 up_read(&inode->i_alloc_sem);
1830 ocfs2_rw_unlock(inode, rw_level);
1836 const struct inode_operations ocfs2_file_iops = {
1837 .setattr = ocfs2_setattr,
1838 .getattr = ocfs2_getattr,
1839 .permission = ocfs2_permission,
1842 const struct inode_operations ocfs2_special_file_iops = {
1843 .setattr = ocfs2_setattr,
1844 .getattr = ocfs2_getattr,
1845 .permission = ocfs2_permission,
1848 const struct file_operations ocfs2_fops = {
1849 .read = do_sync_read,
1850 .write = do_sync_write,
1852 .fsync = ocfs2_sync_file,
1853 .release = ocfs2_file_release,
1854 .open = ocfs2_file_open,
1855 .aio_read = ocfs2_file_aio_read,
1856 .aio_write = ocfs2_file_aio_write,
1857 .ioctl = ocfs2_ioctl,
1858 #ifdef CONFIG_COMPAT
1859 .compat_ioctl = ocfs2_compat_ioctl,
1861 .splice_read = ocfs2_file_splice_read,
1862 .splice_write = ocfs2_file_splice_write,
1865 const struct file_operations ocfs2_dops = {
1866 .read = generic_read_dir,
1867 .readdir = ocfs2_readdir,
1868 .fsync = ocfs2_sync_file,
1869 .ioctl = ocfs2_ioctl,
1870 #ifdef CONFIG_COMPAT
1871 .compat_ioctl = ocfs2_compat_ioctl,