1 /* -*- mode: c; c-basic-offset: 8; -*-
2 * vim: noexpandtab sw=8 ts=8 sts=0:
6 * File open, close, extend, truncate
8 * Copyright (C) 2002, 2004 Oracle. All rights reserved.
10 * This program is free software; you can redistribute it and/or
11 * modify it under the terms of the GNU General Public
12 * License as published by the Free Software Foundation; either
13 * version 2 of the License, or (at your option) any later version.
15 * This program is distributed in the hope that it will be useful,
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
18 * General Public License for more details.
20 * You should have received a copy of the GNU General Public
21 * License along with this program; if not, write to the
22 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
23 * Boston, MA 021110-1307, USA.
26 #include <linux/capability.h>
28 #include <linux/types.h>
29 #include <linux/slab.h>
30 #include <linux/highmem.h>
31 #include <linux/pagemap.h>
32 #include <linux/uio.h>
33 #include <linux/sched.h>
34 #include <linux/splice.h>
35 #include <linux/mount.h>
36 #include <linux/writeback.h>
37 #include <linux/falloc.h>
39 #define MLOG_MASK_PREFIX ML_INODE
40 #include <cluster/masklog.h>
48 #include "extent_map.h"
59 #include "buffer_head_io.h"
61 static int ocfs2_sync_inode(struct inode *inode)
63 filemap_fdatawrite(inode->i_mapping);
64 return sync_mapping_buffers(inode->i_mapping);
67 static int ocfs2_init_file_private(struct inode *inode, struct file *file)
69 struct ocfs2_file_private *fp;
71 fp = kzalloc(sizeof(struct ocfs2_file_private), GFP_KERNEL);
76 mutex_init(&fp->fp_mutex);
77 ocfs2_file_lock_res_init(&fp->fp_flock, fp);
78 file->private_data = fp;
83 static void ocfs2_free_file_private(struct inode *inode, struct file *file)
85 struct ocfs2_file_private *fp = file->private_data;
86 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
89 ocfs2_simple_drop_lockres(osb, &fp->fp_flock);
90 ocfs2_lock_res_free(&fp->fp_flock);
92 file->private_data = NULL;
96 static int ocfs2_file_open(struct inode *inode, struct file *file)
99 int mode = file->f_flags;
100 struct ocfs2_inode_info *oi = OCFS2_I(inode);
102 mlog_entry("(0x%p, 0x%p, '%.*s')\n", inode, file,
103 file->f_path.dentry->d_name.len, file->f_path.dentry->d_name.name);
105 spin_lock(&oi->ip_lock);
107 /* Check that the inode hasn't been wiped from disk by another
108 * node. If it hasn't then we're safe as long as we hold the
109 * spin lock until our increment of open count. */
110 if (OCFS2_I(inode)->ip_flags & OCFS2_INODE_DELETED) {
111 spin_unlock(&oi->ip_lock);
118 oi->ip_flags |= OCFS2_INODE_OPEN_DIRECT;
121 spin_unlock(&oi->ip_lock);
123 status = ocfs2_init_file_private(inode, file);
126 * We want to set open count back if we're failing the
129 spin_lock(&oi->ip_lock);
131 spin_unlock(&oi->ip_lock);
139 static int ocfs2_file_release(struct inode *inode, struct file *file)
141 struct ocfs2_inode_info *oi = OCFS2_I(inode);
143 mlog_entry("(0x%p, 0x%p, '%.*s')\n", inode, file,
144 file->f_path.dentry->d_name.len,
145 file->f_path.dentry->d_name.name);
147 spin_lock(&oi->ip_lock);
148 if (!--oi->ip_open_count)
149 oi->ip_flags &= ~OCFS2_INODE_OPEN_DIRECT;
150 spin_unlock(&oi->ip_lock);
152 ocfs2_free_file_private(inode, file);
159 static int ocfs2_dir_open(struct inode *inode, struct file *file)
161 return ocfs2_init_file_private(inode, file);
164 static int ocfs2_dir_release(struct inode *inode, struct file *file)
166 ocfs2_free_file_private(inode, file);
170 static int ocfs2_sync_file(struct file *file,
171 struct dentry *dentry,
176 struct inode *inode = dentry->d_inode;
177 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
179 mlog_entry("(0x%p, 0x%p, %d, '%.*s')\n", file, dentry, datasync,
180 dentry->d_name.len, dentry->d_name.name);
182 err = ocfs2_sync_inode(dentry->d_inode);
186 journal = osb->journal->j_journal;
187 err = journal_force_commit(journal);
192 return (err < 0) ? -EIO : 0;
195 int ocfs2_should_update_atime(struct inode *inode,
196 struct vfsmount *vfsmnt)
199 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
201 if (ocfs2_is_hard_readonly(osb) || ocfs2_is_soft_readonly(osb))
204 if ((inode->i_flags & S_NOATIME) ||
205 ((inode->i_sb->s_flags & MS_NODIRATIME) && S_ISDIR(inode->i_mode)))
209 * We can be called with no vfsmnt structure - NFSD will
212 * Note that our action here is different than touch_atime() -
213 * if we can't tell whether this is a noatime mount, then we
214 * don't know whether to trust the value of s_atime_quantum.
219 if ((vfsmnt->mnt_flags & MNT_NOATIME) ||
220 ((vfsmnt->mnt_flags & MNT_NODIRATIME) && S_ISDIR(inode->i_mode)))
223 if (vfsmnt->mnt_flags & MNT_RELATIME) {
224 if ((timespec_compare(&inode->i_atime, &inode->i_mtime) <= 0) ||
225 (timespec_compare(&inode->i_atime, &inode->i_ctime) <= 0))
232 if ((now.tv_sec - inode->i_atime.tv_sec <= osb->s_atime_quantum))
238 int ocfs2_update_inode_atime(struct inode *inode,
239 struct buffer_head *bh)
242 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
244 struct ocfs2_dinode *di = (struct ocfs2_dinode *) bh->b_data;
248 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
249 if (handle == NULL) {
255 ret = ocfs2_journal_access(handle, inode, bh,
256 OCFS2_JOURNAL_ACCESS_WRITE);
263 * Don't use ocfs2_mark_inode_dirty() here as we don't always
264 * have i_mutex to guard against concurrent changes to other
267 inode->i_atime = CURRENT_TIME;
268 di->i_atime = cpu_to_le64(inode->i_atime.tv_sec);
269 di->i_atime_nsec = cpu_to_le32(inode->i_atime.tv_nsec);
271 ret = ocfs2_journal_dirty(handle, bh);
276 ocfs2_commit_trans(OCFS2_SB(inode->i_sb), handle);
282 static int ocfs2_set_inode_size(handle_t *handle,
284 struct buffer_head *fe_bh,
290 i_size_write(inode, new_i_size);
291 inode->i_blocks = ocfs2_inode_sector_count(inode);
292 inode->i_ctime = inode->i_mtime = CURRENT_TIME;
294 status = ocfs2_mark_inode_dirty(handle, inode, fe_bh);
305 static int ocfs2_simple_size_update(struct inode *inode,
306 struct buffer_head *di_bh,
310 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
311 handle_t *handle = NULL;
313 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
314 if (handle == NULL) {
320 ret = ocfs2_set_inode_size(handle, inode, di_bh,
325 ocfs2_commit_trans(osb, handle);
330 static int ocfs2_orphan_for_truncate(struct ocfs2_super *osb,
332 struct buffer_head *fe_bh,
337 struct ocfs2_dinode *di;
342 /* TODO: This needs to actually orphan the inode in this
345 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
346 if (IS_ERR(handle)) {
347 status = PTR_ERR(handle);
352 status = ocfs2_journal_access(handle, inode, fe_bh,
353 OCFS2_JOURNAL_ACCESS_WRITE);
360 * Do this before setting i_size.
362 cluster_bytes = ocfs2_align_bytes_to_clusters(inode->i_sb, new_i_size);
363 status = ocfs2_zero_range_for_truncate(inode, handle, new_i_size,
370 i_size_write(inode, new_i_size);
371 inode->i_ctime = inode->i_mtime = CURRENT_TIME;
373 di = (struct ocfs2_dinode *) fe_bh->b_data;
374 di->i_size = cpu_to_le64(new_i_size);
375 di->i_ctime = di->i_mtime = cpu_to_le64(inode->i_ctime.tv_sec);
376 di->i_ctime_nsec = di->i_mtime_nsec = cpu_to_le32(inode->i_ctime.tv_nsec);
378 status = ocfs2_journal_dirty(handle, fe_bh);
383 ocfs2_commit_trans(osb, handle);
390 static int ocfs2_truncate_file(struct inode *inode,
391 struct buffer_head *di_bh,
395 struct ocfs2_dinode *fe = NULL;
396 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
397 struct ocfs2_truncate_context *tc = NULL;
399 mlog_entry("(inode = %llu, new_i_size = %llu\n",
400 (unsigned long long)OCFS2_I(inode)->ip_blkno,
401 (unsigned long long)new_i_size);
403 fe = (struct ocfs2_dinode *) di_bh->b_data;
404 if (!OCFS2_IS_VALID_DINODE(fe)) {
405 OCFS2_RO_ON_INVALID_DINODE(inode->i_sb, fe);
410 mlog_bug_on_msg(le64_to_cpu(fe->i_size) != i_size_read(inode),
411 "Inode %llu, inode i_size = %lld != di "
412 "i_size = %llu, i_flags = 0x%x\n",
413 (unsigned long long)OCFS2_I(inode)->ip_blkno,
415 (unsigned long long)le64_to_cpu(fe->i_size),
416 le32_to_cpu(fe->i_flags));
418 if (new_i_size > le64_to_cpu(fe->i_size)) {
419 mlog(0, "asked to truncate file with size (%llu) to size (%llu)!\n",
420 (unsigned long long)le64_to_cpu(fe->i_size),
421 (unsigned long long)new_i_size);
427 mlog(0, "inode %llu, i_size = %llu, new_i_size = %llu\n",
428 (unsigned long long)le64_to_cpu(fe->i_blkno),
429 (unsigned long long)le64_to_cpu(fe->i_size),
430 (unsigned long long)new_i_size);
432 /* lets handle the simple truncate cases before doing any more
433 * cluster locking. */
434 if (new_i_size == le64_to_cpu(fe->i_size))
437 down_write(&OCFS2_I(inode)->ip_alloc_sem);
440 * The inode lock forced other nodes to sync and drop their
441 * pages, which (correctly) happens even if we have a truncate
442 * without allocation change - ocfs2 cluster sizes can be much
443 * greater than page size, so we have to truncate them
446 unmap_mapping_range(inode->i_mapping, new_i_size + PAGE_SIZE - 1, 0, 1);
447 truncate_inode_pages(inode->i_mapping, new_i_size);
449 if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
450 status = ocfs2_truncate_inline(inode, di_bh, new_i_size,
451 i_size_read(inode), 1);
455 goto bail_unlock_sem;
458 /* alright, we're going to need to do a full blown alloc size
459 * change. Orphan the inode so that recovery can complete the
460 * truncate if necessary. This does the task of marking
462 status = ocfs2_orphan_for_truncate(osb, inode, di_bh, new_i_size);
465 goto bail_unlock_sem;
468 status = ocfs2_prepare_truncate(osb, inode, di_bh, &tc);
471 goto bail_unlock_sem;
474 status = ocfs2_commit_truncate(osb, inode, di_bh, tc);
477 goto bail_unlock_sem;
480 /* TODO: orphan dir cleanup here. */
482 up_write(&OCFS2_I(inode)->ip_alloc_sem);
491 * extend allocation only here.
492 * we'll update all the disk stuff, and oip->alloc_size
494 * expect stuff to be locked, a transaction started and enough data /
495 * metadata reservations in the contexts.
497 * Will return -EAGAIN, and a reason if a restart is needed.
498 * If passed in, *reason will always be set, even in error.
500 int ocfs2_do_extend_allocation(struct ocfs2_super *osb,
505 struct buffer_head *fe_bh,
507 struct ocfs2_alloc_context *data_ac,
508 struct ocfs2_alloc_context *meta_ac,
509 enum ocfs2_alloc_restarted *reason_ret)
513 struct ocfs2_dinode *fe = (struct ocfs2_dinode *) fe_bh->b_data;
514 enum ocfs2_alloc_restarted reason = RESTART_NONE;
515 u32 bit_off, num_bits;
519 BUG_ON(!clusters_to_add);
522 flags = OCFS2_EXT_UNWRITTEN;
524 free_extents = ocfs2_num_free_extents(osb, inode, fe_bh);
525 if (free_extents < 0) {
526 status = free_extents;
531 /* there are two cases which could cause us to EAGAIN in the
532 * we-need-more-metadata case:
533 * 1) we haven't reserved *any*
534 * 2) we are so fragmented, we've needed to add metadata too
536 if (!free_extents && !meta_ac) {
537 mlog(0, "we haven't reserved any metadata!\n");
539 reason = RESTART_META;
541 } else if ((!free_extents)
542 && (ocfs2_alloc_context_bits_left(meta_ac)
543 < ocfs2_extend_meta_needed(&fe->id2.i_list))) {
544 mlog(0, "filesystem is really fragmented...\n");
546 reason = RESTART_META;
550 status = __ocfs2_claim_clusters(osb, handle, data_ac, 1,
551 clusters_to_add, &bit_off, &num_bits);
553 if (status != -ENOSPC)
558 BUG_ON(num_bits > clusters_to_add);
560 /* reserve our write early -- insert_extent may update the inode */
561 status = ocfs2_journal_access(handle, inode, fe_bh,
562 OCFS2_JOURNAL_ACCESS_WRITE);
568 block = ocfs2_clusters_to_blocks(osb->sb, bit_off);
569 mlog(0, "Allocating %u clusters at block %u for inode %llu\n",
570 num_bits, bit_off, (unsigned long long)OCFS2_I(inode)->ip_blkno);
571 status = ocfs2_insert_extent(osb, handle, inode, fe_bh,
572 *logical_offset, block, num_bits,
579 status = ocfs2_journal_dirty(handle, fe_bh);
585 clusters_to_add -= num_bits;
586 *logical_offset += num_bits;
588 if (clusters_to_add) {
589 mlog(0, "need to alloc once more, clusters = %u, wanted = "
590 "%u\n", fe->i_clusters, clusters_to_add);
592 reason = RESTART_TRANS;
598 *reason_ret = reason;
603 * For a given allocation, determine which allocators will need to be
604 * accessed, and lock them, reserving the appropriate number of bits.
606 * Sparse file systems call this from ocfs2_write_begin_nolock()
607 * and ocfs2_allocate_unwritten_extents().
609 * File systems which don't support holes call this from
610 * ocfs2_extend_allocation().
612 int ocfs2_lock_allocators(struct inode *inode, struct buffer_head *di_bh,
613 u32 clusters_to_add, u32 extents_to_split,
614 struct ocfs2_alloc_context **data_ac,
615 struct ocfs2_alloc_context **meta_ac)
617 int ret = 0, num_free_extents;
618 unsigned int max_recs_needed = clusters_to_add + 2 * extents_to_split;
619 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
620 struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
626 BUG_ON(clusters_to_add != 0 && data_ac == NULL);
628 mlog(0, "extend inode %llu, i_size = %lld, di->i_clusters = %u, "
629 "clusters_to_add = %u, extents_to_split = %u\n",
630 (unsigned long long)OCFS2_I(inode)->ip_blkno, (long long)i_size_read(inode),
631 le32_to_cpu(di->i_clusters), clusters_to_add, extents_to_split);
633 num_free_extents = ocfs2_num_free_extents(osb, inode, di_bh);
634 if (num_free_extents < 0) {
635 ret = num_free_extents;
641 * Sparse allocation file systems need to be more conservative
642 * with reserving room for expansion - the actual allocation
643 * happens while we've got a journal handle open so re-taking
644 * a cluster lock (because we ran out of room for another
645 * extent) will violate ordering rules.
647 * Most of the time we'll only be seeing this 1 cluster at a time
650 * Always lock for any unwritten extents - we might want to
651 * add blocks during a split.
653 if (!num_free_extents ||
654 (ocfs2_sparse_alloc(osb) && num_free_extents < max_recs_needed)) {
655 ret = ocfs2_reserve_new_metadata(osb, &di->id2.i_list, meta_ac);
663 if (clusters_to_add == 0)
666 ret = ocfs2_reserve_clusters(osb, clusters_to_add, data_ac);
676 ocfs2_free_alloc_context(*meta_ac);
681 * We cannot have an error and a non null *data_ac.
688 static int __ocfs2_extend_allocation(struct inode *inode, u32 logical_start,
689 u32 clusters_to_add, int mark_unwritten)
692 int restart_func = 0;
695 struct buffer_head *bh = NULL;
696 struct ocfs2_dinode *fe = NULL;
697 handle_t *handle = NULL;
698 struct ocfs2_alloc_context *data_ac = NULL;
699 struct ocfs2_alloc_context *meta_ac = NULL;
700 enum ocfs2_alloc_restarted why;
701 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
703 mlog_entry("(clusters_to_add = %u)\n", clusters_to_add);
706 * This function only exists for file systems which don't
709 BUG_ON(mark_unwritten && !ocfs2_sparse_alloc(osb));
711 status = ocfs2_read_block(osb, OCFS2_I(inode)->ip_blkno, &bh,
712 OCFS2_BH_CACHED, inode);
718 fe = (struct ocfs2_dinode *) bh->b_data;
719 if (!OCFS2_IS_VALID_DINODE(fe)) {
720 OCFS2_RO_ON_INVALID_DINODE(inode->i_sb, fe);
726 BUG_ON(le32_to_cpu(fe->i_clusters) != OCFS2_I(inode)->ip_clusters);
728 status = ocfs2_lock_allocators(inode, bh, clusters_to_add, 0, &data_ac,
735 credits = ocfs2_calc_extend_credits(osb->sb, &fe->id2.i_list,
737 handle = ocfs2_start_trans(osb, credits);
738 if (IS_ERR(handle)) {
739 status = PTR_ERR(handle);
745 restarted_transaction:
746 /* reserve a write to the file entry early on - that we if we
747 * run out of credits in the allocation path, we can still
749 status = ocfs2_journal_access(handle, inode, bh,
750 OCFS2_JOURNAL_ACCESS_WRITE);
756 prev_clusters = OCFS2_I(inode)->ip_clusters;
758 status = ocfs2_do_extend_allocation(osb,
768 if ((status < 0) && (status != -EAGAIN)) {
769 if (status != -ENOSPC)
774 status = ocfs2_journal_dirty(handle, bh);
780 spin_lock(&OCFS2_I(inode)->ip_lock);
781 clusters_to_add -= (OCFS2_I(inode)->ip_clusters - prev_clusters);
782 spin_unlock(&OCFS2_I(inode)->ip_lock);
784 if (why != RESTART_NONE && clusters_to_add) {
785 if (why == RESTART_META) {
786 mlog(0, "restarting function.\n");
789 BUG_ON(why != RESTART_TRANS);
791 mlog(0, "restarting transaction.\n");
792 /* TODO: This can be more intelligent. */
793 credits = ocfs2_calc_extend_credits(osb->sb,
796 status = ocfs2_extend_trans(handle, credits);
798 /* handle still has to be committed at
804 goto restarted_transaction;
808 mlog(0, "fe: i_clusters = %u, i_size=%llu\n",
809 le32_to_cpu(fe->i_clusters),
810 (unsigned long long)le64_to_cpu(fe->i_size));
811 mlog(0, "inode: ip_clusters=%u, i_size=%lld\n",
812 OCFS2_I(inode)->ip_clusters, (long long)i_size_read(inode));
816 ocfs2_commit_trans(osb, handle);
820 ocfs2_free_alloc_context(data_ac);
824 ocfs2_free_alloc_context(meta_ac);
827 if ((!status) && restart_func) {
840 /* Some parts of this taken from generic_cont_expand, which turned out
841 * to be too fragile to do exactly what we need without us having to
842 * worry about recursive locking in ->prepare_write() and
843 * ->commit_write(). */
844 static int ocfs2_write_zero_page(struct inode *inode,
847 struct address_space *mapping = inode->i_mapping;
851 handle_t *handle = NULL;
854 offset = (size & (PAGE_CACHE_SIZE-1)); /* Within page */
855 /* ugh. in prepare/commit_write, if from==to==start of block, we
856 ** skip the prepare. make sure we never send an offset for the start
859 if ((offset & (inode->i_sb->s_blocksize - 1)) == 0) {
862 index = size >> PAGE_CACHE_SHIFT;
864 page = grab_cache_page(mapping, index);
871 ret = ocfs2_prepare_write_nolock(inode, page, offset, offset);
877 if (ocfs2_should_order_data(inode)) {
878 handle = ocfs2_start_walk_page_trans(inode, page, offset,
880 if (IS_ERR(handle)) {
881 ret = PTR_ERR(handle);
887 /* must not update i_size! */
888 ret = block_commit_write(page, offset, offset);
895 ocfs2_commit_trans(OCFS2_SB(inode->i_sb), handle);
898 page_cache_release(page);
903 static int ocfs2_zero_extend(struct inode *inode,
908 struct super_block *sb = inode->i_sb;
910 start_off = ocfs2_align_bytes_to_blocks(sb, i_size_read(inode));
911 while (start_off < zero_to_size) {
912 ret = ocfs2_write_zero_page(inode, start_off);
918 start_off += sb->s_blocksize;
921 * Very large extends have the potential to lock up
922 * the cpu for extended periods of time.
931 int ocfs2_extend_no_holes(struct inode *inode, u64 new_i_size, u64 zero_to)
935 struct ocfs2_inode_info *oi = OCFS2_I(inode);
937 clusters_to_add = ocfs2_clusters_for_bytes(inode->i_sb, new_i_size);
938 if (clusters_to_add < oi->ip_clusters)
941 clusters_to_add -= oi->ip_clusters;
943 if (clusters_to_add) {
944 ret = __ocfs2_extend_allocation(inode, oi->ip_clusters,
953 * Call this even if we don't add any clusters to the tree. We
954 * still need to zero the area between the old i_size and the
957 ret = ocfs2_zero_extend(inode, zero_to);
965 static int ocfs2_extend_file(struct inode *inode,
966 struct buffer_head *di_bh,
970 struct ocfs2_inode_info *oi = OCFS2_I(inode);
974 /* setattr sometimes calls us like this. */
978 if (i_size_read(inode) == new_i_size)
980 BUG_ON(new_i_size < i_size_read(inode));
983 * Fall through for converting inline data, even if the fs
984 * supports sparse files.
986 * The check for inline data here is legal - nobody can add
987 * the feature since we have i_mutex. We must check it again
988 * after acquiring ip_alloc_sem though, as paths like mmap
989 * might have raced us to converting the inode to extents.
991 if (!(oi->ip_dyn_features & OCFS2_INLINE_DATA_FL)
992 && ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)))
993 goto out_update_size;
996 * The alloc sem blocks people in read/write from reading our
997 * allocation until we're done changing it. We depend on
998 * i_mutex to block other extend/truncate calls while we're
1001 down_write(&oi->ip_alloc_sem);
1003 if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1005 * We can optimize small extends by keeping the inodes
1008 if (ocfs2_size_fits_inline_data(di_bh, new_i_size)) {
1009 up_write(&oi->ip_alloc_sem);
1010 goto out_update_size;
1013 ret = ocfs2_convert_inline_data_to_extents(inode, di_bh);
1015 up_write(&oi->ip_alloc_sem);
1022 if (!ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)))
1023 ret = ocfs2_extend_no_holes(inode, new_i_size, new_i_size);
1025 up_write(&oi->ip_alloc_sem);
1033 ret = ocfs2_simple_size_update(inode, di_bh, new_i_size);
1041 int ocfs2_setattr(struct dentry *dentry, struct iattr *attr)
1043 int status = 0, size_change;
1044 struct inode *inode = dentry->d_inode;
1045 struct super_block *sb = inode->i_sb;
1046 struct ocfs2_super *osb = OCFS2_SB(sb);
1047 struct buffer_head *bh = NULL;
1048 handle_t *handle = NULL;
1050 mlog_entry("(0x%p, '%.*s')\n", dentry,
1051 dentry->d_name.len, dentry->d_name.name);
1053 /* ensuring we don't even attempt to truncate a symlink */
1054 if (S_ISLNK(inode->i_mode))
1055 attr->ia_valid &= ~ATTR_SIZE;
1057 if (attr->ia_valid & ATTR_MODE)
1058 mlog(0, "mode change: %d\n", attr->ia_mode);
1059 if (attr->ia_valid & ATTR_UID)
1060 mlog(0, "uid change: %d\n", attr->ia_uid);
1061 if (attr->ia_valid & ATTR_GID)
1062 mlog(0, "gid change: %d\n", attr->ia_gid);
1063 if (attr->ia_valid & ATTR_SIZE)
1064 mlog(0, "size change...\n");
1065 if (attr->ia_valid & (ATTR_ATIME | ATTR_MTIME | ATTR_CTIME))
1066 mlog(0, "time change...\n");
1068 #define OCFS2_VALID_ATTRS (ATTR_ATIME | ATTR_MTIME | ATTR_CTIME | ATTR_SIZE \
1069 | ATTR_GID | ATTR_UID | ATTR_MODE)
1070 if (!(attr->ia_valid & OCFS2_VALID_ATTRS)) {
1071 mlog(0, "can't handle attrs: 0x%x\n", attr->ia_valid);
1075 status = inode_change_ok(inode, attr);
1079 size_change = S_ISREG(inode->i_mode) && attr->ia_valid & ATTR_SIZE;
1081 status = ocfs2_rw_lock(inode, 1);
1088 status = ocfs2_inode_lock(inode, &bh, 1);
1090 if (status != -ENOENT)
1092 goto bail_unlock_rw;
1095 if (size_change && attr->ia_size != i_size_read(inode)) {
1096 if (attr->ia_size > sb->s_maxbytes) {
1101 if (i_size_read(inode) > attr->ia_size)
1102 status = ocfs2_truncate_file(inode, bh, attr->ia_size);
1104 status = ocfs2_extend_file(inode, bh, attr->ia_size);
1106 if (status != -ENOSPC)
1113 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1114 if (IS_ERR(handle)) {
1115 status = PTR_ERR(handle);
1121 * This will intentionally not wind up calling vmtruncate(),
1122 * since all the work for a size change has been done above.
1123 * Otherwise, we could get into problems with truncate as
1124 * ip_alloc_sem is used there to protect against i_size
1127 status = inode_setattr(inode, attr);
1133 status = ocfs2_mark_inode_dirty(handle, inode, bh);
1138 ocfs2_commit_trans(osb, handle);
1140 ocfs2_inode_unlock(inode, 1);
1143 ocfs2_rw_unlock(inode, 1);
1152 int ocfs2_getattr(struct vfsmount *mnt,
1153 struct dentry *dentry,
1156 struct inode *inode = dentry->d_inode;
1157 struct super_block *sb = dentry->d_inode->i_sb;
1158 struct ocfs2_super *osb = sb->s_fs_info;
1163 err = ocfs2_inode_revalidate(dentry);
1170 generic_fillattr(inode, stat);
1172 /* We set the blksize from the cluster size for performance */
1173 stat->blksize = osb->s_clustersize;
1181 int ocfs2_permission(struct inode *inode, int mask)
1187 ret = ocfs2_inode_lock(inode, NULL, 0);
1194 ret = generic_permission(inode, mask, NULL);
1196 ocfs2_inode_unlock(inode, 0);
1202 static int __ocfs2_write_remove_suid(struct inode *inode,
1203 struct buffer_head *bh)
1207 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1208 struct ocfs2_dinode *di;
1210 mlog_entry("(Inode %llu, mode 0%o)\n",
1211 (unsigned long long)OCFS2_I(inode)->ip_blkno, inode->i_mode);
1213 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1214 if (handle == NULL) {
1220 ret = ocfs2_journal_access(handle, inode, bh,
1221 OCFS2_JOURNAL_ACCESS_WRITE);
1227 inode->i_mode &= ~S_ISUID;
1228 if ((inode->i_mode & S_ISGID) && (inode->i_mode & S_IXGRP))
1229 inode->i_mode &= ~S_ISGID;
1231 di = (struct ocfs2_dinode *) bh->b_data;
1232 di->i_mode = cpu_to_le16(inode->i_mode);
1234 ret = ocfs2_journal_dirty(handle, bh);
1239 ocfs2_commit_trans(osb, handle);
1246 * Will look for holes and unwritten extents in the range starting at
1247 * pos for count bytes (inclusive).
1249 static int ocfs2_check_range_for_holes(struct inode *inode, loff_t pos,
1253 unsigned int extent_flags;
1254 u32 cpos, clusters, extent_len, phys_cpos;
1255 struct super_block *sb = inode->i_sb;
1257 cpos = pos >> OCFS2_SB(sb)->s_clustersize_bits;
1258 clusters = ocfs2_clusters_for_bytes(sb, pos + count) - cpos;
1261 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos, &extent_len,
1268 if (phys_cpos == 0 || (extent_flags & OCFS2_EXT_UNWRITTEN)) {
1273 if (extent_len > clusters)
1274 extent_len = clusters;
1276 clusters -= extent_len;
1283 static int ocfs2_write_remove_suid(struct inode *inode)
1286 struct buffer_head *bh = NULL;
1287 struct ocfs2_inode_info *oi = OCFS2_I(inode);
1289 ret = ocfs2_read_block(OCFS2_SB(inode->i_sb),
1290 oi->ip_blkno, &bh, OCFS2_BH_CACHED, inode);
1296 ret = __ocfs2_write_remove_suid(inode, bh);
1303 * Allocate enough extents to cover the region starting at byte offset
1304 * start for len bytes. Existing extents are skipped, any extents
1305 * added are marked as "unwritten".
1307 static int ocfs2_allocate_unwritten_extents(struct inode *inode,
1311 u32 cpos, phys_cpos, clusters, alloc_size;
1312 u64 end = start + len;
1313 struct buffer_head *di_bh = NULL;
1315 if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1316 ret = ocfs2_read_block(OCFS2_SB(inode->i_sb),
1317 OCFS2_I(inode)->ip_blkno, &di_bh,
1318 OCFS2_BH_CACHED, inode);
1325 * Nothing to do if the requested reservation range
1326 * fits within the inode.
1328 if (ocfs2_size_fits_inline_data(di_bh, end))
1331 ret = ocfs2_convert_inline_data_to_extents(inode, di_bh);
1339 * We consider both start and len to be inclusive.
1341 cpos = start >> OCFS2_SB(inode->i_sb)->s_clustersize_bits;
1342 clusters = ocfs2_clusters_for_bytes(inode->i_sb, start + len);
1346 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos,
1354 * Hole or existing extent len can be arbitrary, so
1355 * cap it to our own allocation request.
1357 if (alloc_size > clusters)
1358 alloc_size = clusters;
1362 * We already have an allocation at this
1363 * region so we can safely skip it.
1368 ret = __ocfs2_extend_allocation(inode, cpos, alloc_size, 1);
1377 clusters -= alloc_size;
1387 static int __ocfs2_remove_inode_range(struct inode *inode,
1388 struct buffer_head *di_bh,
1389 u32 cpos, u32 phys_cpos, u32 len,
1390 struct ocfs2_cached_dealloc_ctxt *dealloc)
1393 u64 phys_blkno = ocfs2_clusters_to_blocks(inode->i_sb, phys_cpos);
1394 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1395 struct inode *tl_inode = osb->osb_tl_inode;
1397 struct ocfs2_alloc_context *meta_ac = NULL;
1398 struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
1400 ret = ocfs2_lock_allocators(inode, di_bh, 0, 1, NULL, &meta_ac);
1406 mutex_lock(&tl_inode->i_mutex);
1408 if (ocfs2_truncate_log_needs_flush(osb)) {
1409 ret = __ocfs2_flush_truncate_log(osb);
1416 handle = ocfs2_start_trans(osb, OCFS2_REMOVE_EXTENT_CREDITS);
1417 if (handle == NULL) {
1423 ret = ocfs2_journal_access(handle, inode, di_bh,
1424 OCFS2_JOURNAL_ACCESS_WRITE);
1430 ret = ocfs2_remove_extent(inode, di_bh, cpos, len, handle, meta_ac,
1437 OCFS2_I(inode)->ip_clusters -= len;
1438 di->i_clusters = cpu_to_le32(OCFS2_I(inode)->ip_clusters);
1440 ret = ocfs2_journal_dirty(handle, di_bh);
1446 ret = ocfs2_truncate_log_append(osb, handle, phys_blkno, len);
1451 ocfs2_commit_trans(osb, handle);
1453 mutex_unlock(&tl_inode->i_mutex);
1456 ocfs2_free_alloc_context(meta_ac);
1462 * Truncate a byte range, avoiding pages within partial clusters. This
1463 * preserves those pages for the zeroing code to write to.
1465 static void ocfs2_truncate_cluster_pages(struct inode *inode, u64 byte_start,
1468 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1470 struct address_space *mapping = inode->i_mapping;
1472 start = (loff_t)ocfs2_align_bytes_to_clusters(inode->i_sb, byte_start);
1473 end = byte_start + byte_len;
1474 end = end & ~(osb->s_clustersize - 1);
1477 unmap_mapping_range(mapping, start, end - start, 0);
1478 truncate_inode_pages_range(mapping, start, end - 1);
1482 static int ocfs2_zero_partial_clusters(struct inode *inode,
1486 u64 tmpend, end = start + len;
1487 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1488 unsigned int csize = osb->s_clustersize;
1492 * The "start" and "end" values are NOT necessarily part of
1493 * the range whose allocation is being deleted. Rather, this
1494 * is what the user passed in with the request. We must zero
1495 * partial clusters here. There's no need to worry about
1496 * physical allocation - the zeroing code knows to skip holes.
1498 mlog(0, "byte start: %llu, end: %llu\n",
1499 (unsigned long long)start, (unsigned long long)end);
1502 * If both edges are on a cluster boundary then there's no
1503 * zeroing required as the region is part of the allocation to
1506 if ((start & (csize - 1)) == 0 && (end & (csize - 1)) == 0)
1509 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1510 if (handle == NULL) {
1517 * We want to get the byte offset of the end of the 1st cluster.
1519 tmpend = (u64)osb->s_clustersize + (start & ~(osb->s_clustersize - 1));
1523 mlog(0, "1st range: start: %llu, tmpend: %llu\n",
1524 (unsigned long long)start, (unsigned long long)tmpend);
1526 ret = ocfs2_zero_range_for_truncate(inode, handle, start, tmpend);
1532 * This may make start and end equal, but the zeroing
1533 * code will skip any work in that case so there's no
1534 * need to catch it up here.
1536 start = end & ~(osb->s_clustersize - 1);
1538 mlog(0, "2nd range: start: %llu, end: %llu\n",
1539 (unsigned long long)start, (unsigned long long)end);
1541 ret = ocfs2_zero_range_for_truncate(inode, handle, start, end);
1546 ocfs2_commit_trans(osb, handle);
1551 static int ocfs2_remove_inode_range(struct inode *inode,
1552 struct buffer_head *di_bh, u64 byte_start,
1556 u32 trunc_start, trunc_len, cpos, phys_cpos, alloc_size;
1557 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1558 struct ocfs2_cached_dealloc_ctxt dealloc;
1559 struct address_space *mapping = inode->i_mapping;
1561 ocfs2_init_dealloc_ctxt(&dealloc);
1566 if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1567 ret = ocfs2_truncate_inline(inode, di_bh, byte_start,
1568 byte_start + byte_len, 0);
1574 * There's no need to get fancy with the page cache
1575 * truncate of an inline-data inode. We're talking
1576 * about less than a page here, which will be cached
1577 * in the dinode buffer anyway.
1579 unmap_mapping_range(mapping, 0, 0, 0);
1580 truncate_inode_pages(mapping, 0);
1584 trunc_start = ocfs2_clusters_for_bytes(osb->sb, byte_start);
1585 trunc_len = (byte_start + byte_len) >> osb->s_clustersize_bits;
1586 if (trunc_len >= trunc_start)
1587 trunc_len -= trunc_start;
1591 mlog(0, "Inode: %llu, start: %llu, len: %llu, cstart: %u, clen: %u\n",
1592 (unsigned long long)OCFS2_I(inode)->ip_blkno,
1593 (unsigned long long)byte_start,
1594 (unsigned long long)byte_len, trunc_start, trunc_len);
1596 ret = ocfs2_zero_partial_clusters(inode, byte_start, byte_len);
1604 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos,
1611 if (alloc_size > trunc_len)
1612 alloc_size = trunc_len;
1614 /* Only do work for non-holes */
1615 if (phys_cpos != 0) {
1616 ret = __ocfs2_remove_inode_range(inode, di_bh, cpos,
1617 phys_cpos, alloc_size,
1626 trunc_len -= alloc_size;
1629 ocfs2_truncate_cluster_pages(inode, byte_start, byte_len);
1632 ocfs2_schedule_truncate_log_flush(osb, 1);
1633 ocfs2_run_deallocs(osb, &dealloc);
1639 * Parts of this function taken from xfs_change_file_space()
1641 static int __ocfs2_change_file_space(struct file *file, struct inode *inode,
1642 loff_t f_pos, unsigned int cmd,
1643 struct ocfs2_space_resv *sr,
1649 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1650 struct buffer_head *di_bh = NULL;
1652 unsigned long long max_off = inode->i_sb->s_maxbytes;
1654 if (ocfs2_is_hard_readonly(osb) || ocfs2_is_soft_readonly(osb))
1657 mutex_lock(&inode->i_mutex);
1660 * This prevents concurrent writes on other nodes
1662 ret = ocfs2_rw_lock(inode, 1);
1668 ret = ocfs2_inode_lock(inode, &di_bh, 1);
1674 if (inode->i_flags & (S_IMMUTABLE|S_APPEND)) {
1676 goto out_inode_unlock;
1679 switch (sr->l_whence) {
1680 case 0: /*SEEK_SET*/
1682 case 1: /*SEEK_CUR*/
1683 sr->l_start += f_pos;
1685 case 2: /*SEEK_END*/
1686 sr->l_start += i_size_read(inode);
1690 goto out_inode_unlock;
1694 llen = sr->l_len > 0 ? sr->l_len - 1 : sr->l_len;
1697 || sr->l_start > max_off
1698 || (sr->l_start + llen) < 0
1699 || (sr->l_start + llen) > max_off) {
1701 goto out_inode_unlock;
1703 size = sr->l_start + sr->l_len;
1705 if (cmd == OCFS2_IOC_RESVSP || cmd == OCFS2_IOC_RESVSP64) {
1706 if (sr->l_len <= 0) {
1708 goto out_inode_unlock;
1712 if (file && should_remove_suid(file->f_path.dentry)) {
1713 ret = __ocfs2_write_remove_suid(inode, di_bh);
1716 goto out_inode_unlock;
1720 down_write(&OCFS2_I(inode)->ip_alloc_sem);
1722 case OCFS2_IOC_RESVSP:
1723 case OCFS2_IOC_RESVSP64:
1725 * This takes unsigned offsets, but the signed ones we
1726 * pass have been checked against overflow above.
1728 ret = ocfs2_allocate_unwritten_extents(inode, sr->l_start,
1731 case OCFS2_IOC_UNRESVSP:
1732 case OCFS2_IOC_UNRESVSP64:
1733 ret = ocfs2_remove_inode_range(inode, di_bh, sr->l_start,
1739 up_write(&OCFS2_I(inode)->ip_alloc_sem);
1742 goto out_inode_unlock;
1746 * We update c/mtime for these changes
1748 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1749 if (IS_ERR(handle)) {
1750 ret = PTR_ERR(handle);
1752 goto out_inode_unlock;
1755 if (change_size && i_size_read(inode) < size)
1756 i_size_write(inode, size);
1758 inode->i_ctime = inode->i_mtime = CURRENT_TIME;
1759 ret = ocfs2_mark_inode_dirty(handle, inode, di_bh);
1763 ocfs2_commit_trans(osb, handle);
1767 ocfs2_inode_unlock(inode, 1);
1769 ocfs2_rw_unlock(inode, 1);
1772 mutex_unlock(&inode->i_mutex);
1776 int ocfs2_change_file_space(struct file *file, unsigned int cmd,
1777 struct ocfs2_space_resv *sr)
1779 struct inode *inode = file->f_path.dentry->d_inode;
1780 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);;
1782 if ((cmd == OCFS2_IOC_RESVSP || cmd == OCFS2_IOC_RESVSP64) &&
1783 !ocfs2_writes_unwritten_extents(osb))
1785 else if ((cmd == OCFS2_IOC_UNRESVSP || cmd == OCFS2_IOC_UNRESVSP64) &&
1786 !ocfs2_sparse_alloc(osb))
1789 if (!S_ISREG(inode->i_mode))
1792 if (!(file->f_mode & FMODE_WRITE))
1795 return __ocfs2_change_file_space(file, inode, file->f_pos, cmd, sr, 0);
1798 static long ocfs2_fallocate(struct inode *inode, int mode, loff_t offset,
1801 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1802 struct ocfs2_space_resv sr;
1803 int change_size = 1;
1805 if (!ocfs2_writes_unwritten_extents(osb))
1808 if (S_ISDIR(inode->i_mode))
1811 if (mode & FALLOC_FL_KEEP_SIZE)
1815 sr.l_start = (s64)offset;
1816 sr.l_len = (s64)len;
1818 return __ocfs2_change_file_space(NULL, inode, offset,
1819 OCFS2_IOC_RESVSP64, &sr, change_size);
1822 static int ocfs2_prepare_inode_for_write(struct dentry *dentry,
1828 int ret = 0, meta_level = 0;
1829 struct inode *inode = dentry->d_inode;
1830 loff_t saved_pos, end;
1833 * We start with a read level meta lock and only jump to an ex
1834 * if we need to make modifications here.
1837 ret = ocfs2_inode_lock(inode, NULL, meta_level);
1844 /* Clear suid / sgid if necessary. We do this here
1845 * instead of later in the write path because
1846 * remove_suid() calls ->setattr without any hint that
1847 * we may have already done our cluster locking. Since
1848 * ocfs2_setattr() *must* take cluster locks to
1849 * proceeed, this will lead us to recursively lock the
1850 * inode. There's also the dinode i_size state which
1851 * can be lost via setattr during extending writes (we
1852 * set inode->i_size at the end of a write. */
1853 if (should_remove_suid(dentry)) {
1854 if (meta_level == 0) {
1855 ocfs2_inode_unlock(inode, meta_level);
1860 ret = ocfs2_write_remove_suid(inode);
1867 /* work on a copy of ppos until we're sure that we won't have
1868 * to recalculate it due to relocking. */
1870 saved_pos = i_size_read(inode);
1871 mlog(0, "O_APPEND: inode->i_size=%llu\n", saved_pos);
1876 end = saved_pos + count;
1879 * Skip the O_DIRECT checks if we don't need
1882 if (!direct_io || !(*direct_io))
1886 * There's no sane way to do direct writes to an inode
1889 if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1895 * Allowing concurrent direct writes means
1896 * i_size changes wouldn't be synchronized, so
1897 * one node could wind up truncating another
1900 if (end > i_size_read(inode)) {
1906 * We don't fill holes during direct io, so
1907 * check for them here. If any are found, the
1908 * caller will have to retake some cluster
1909 * locks and initiate the io as buffered.
1911 ret = ocfs2_check_range_for_holes(inode, saved_pos, count);
1924 ocfs2_inode_unlock(inode, meta_level);
1930 static ssize_t ocfs2_file_aio_write(struct kiocb *iocb,
1931 const struct iovec *iov,
1932 unsigned long nr_segs,
1935 int ret, direct_io, appending, rw_level, have_alloc_sem = 0;
1937 ssize_t written = 0;
1938 size_t ocount; /* original count */
1939 size_t count; /* after file limit checks */
1940 loff_t old_size, *ppos = &iocb->ki_pos;
1942 struct file *file = iocb->ki_filp;
1943 struct inode *inode = file->f_path.dentry->d_inode;
1944 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1946 mlog_entry("(0x%p, %u, '%.*s')\n", file,
1947 (unsigned int)nr_segs,
1948 file->f_path.dentry->d_name.len,
1949 file->f_path.dentry->d_name.name);
1951 if (iocb->ki_left == 0)
1954 vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
1956 appending = file->f_flags & O_APPEND ? 1 : 0;
1957 direct_io = file->f_flags & O_DIRECT ? 1 : 0;
1959 mutex_lock(&inode->i_mutex);
1962 /* to match setattr's i_mutex -> i_alloc_sem -> rw_lock ordering */
1964 down_read(&inode->i_alloc_sem);
1968 /* concurrent O_DIRECT writes are allowed */
1969 rw_level = !direct_io;
1970 ret = ocfs2_rw_lock(inode, rw_level);
1976 can_do_direct = direct_io;
1977 ret = ocfs2_prepare_inode_for_write(file->f_path.dentry, ppos,
1978 iocb->ki_left, appending,
1986 * We can't complete the direct I/O as requested, fall back to
1989 if (direct_io && !can_do_direct) {
1990 ocfs2_rw_unlock(inode, rw_level);
1991 up_read(&inode->i_alloc_sem);
2001 * To later detect whether a journal commit for sync writes is
2002 * necessary, we sample i_size, and cluster count here.
2004 old_size = i_size_read(inode);
2005 old_clusters = OCFS2_I(inode)->ip_clusters;
2007 /* communicate with ocfs2_dio_end_io */
2008 ocfs2_iocb_set_rw_locked(iocb, rw_level);
2011 ret = generic_segment_checks(iov, &nr_segs, &ocount,
2016 ret = generic_write_checks(file, ppos, &count,
2017 S_ISBLK(inode->i_mode));
2021 written = generic_file_direct_write(iocb, iov, &nr_segs, *ppos,
2022 ppos, count, ocount);
2028 written = generic_file_aio_write_nolock(iocb, iov, nr_segs,
2033 /* buffered aio wouldn't have proper lock coverage today */
2034 BUG_ON(ret == -EIOCBQUEUED && !(file->f_flags & O_DIRECT));
2036 if ((file->f_flags & O_SYNC && !direct_io) || IS_SYNC(inode)) {
2038 * The generic write paths have handled getting data
2039 * to disk, but since we don't make use of the dirty
2040 * inode list, a manual journal commit is necessary
2043 if (old_size != i_size_read(inode) ||
2044 old_clusters != OCFS2_I(inode)->ip_clusters) {
2045 ret = journal_force_commit(osb->journal->j_journal);
2052 * deep in g_f_a_w_n()->ocfs2_direct_IO we pass in a ocfs2_dio_end_io
2053 * function pointer which is called when o_direct io completes so that
2054 * it can unlock our rw lock. (it's the clustered equivalent of
2055 * i_alloc_sem; protects truncate from racing with pending ios).
2056 * Unfortunately there are error cases which call end_io and others
2057 * that don't. so we don't have to unlock the rw_lock if either an
2058 * async dio is going to do it in the future or an end_io after an
2059 * error has already done it.
2061 if (ret == -EIOCBQUEUED || !ocfs2_iocb_is_rw_locked(iocb)) {
2068 ocfs2_rw_unlock(inode, rw_level);
2072 up_read(&inode->i_alloc_sem);
2074 mutex_unlock(&inode->i_mutex);
2077 return written ? written : ret;
2080 static ssize_t ocfs2_file_splice_write(struct pipe_inode_info *pipe,
2087 struct inode *inode = out->f_path.dentry->d_inode;
2089 mlog_entry("(0x%p, 0x%p, %u, '%.*s')\n", out, pipe,
2091 out->f_path.dentry->d_name.len,
2092 out->f_path.dentry->d_name.name);
2094 inode_double_lock(inode, pipe->inode);
2096 ret = ocfs2_rw_lock(inode, 1);
2102 ret = ocfs2_prepare_inode_for_write(out->f_path.dentry, ppos, len, 0,
2109 ret = generic_file_splice_write_nolock(pipe, out, ppos, len, flags);
2112 ocfs2_rw_unlock(inode, 1);
2114 inode_double_unlock(inode, pipe->inode);
2120 static ssize_t ocfs2_file_splice_read(struct file *in,
2122 struct pipe_inode_info *pipe,
2127 struct inode *inode = in->f_path.dentry->d_inode;
2129 mlog_entry("(0x%p, 0x%p, %u, '%.*s')\n", in, pipe,
2131 in->f_path.dentry->d_name.len,
2132 in->f_path.dentry->d_name.name);
2135 * See the comment in ocfs2_file_aio_read()
2137 ret = ocfs2_inode_lock(inode, NULL, 0);
2142 ocfs2_inode_unlock(inode, 0);
2144 ret = generic_file_splice_read(in, ppos, pipe, len, flags);
2151 static ssize_t ocfs2_file_aio_read(struct kiocb *iocb,
2152 const struct iovec *iov,
2153 unsigned long nr_segs,
2156 int ret = 0, rw_level = -1, have_alloc_sem = 0, lock_level = 0;
2157 struct file *filp = iocb->ki_filp;
2158 struct inode *inode = filp->f_path.dentry->d_inode;
2160 mlog_entry("(0x%p, %u, '%.*s')\n", filp,
2161 (unsigned int)nr_segs,
2162 filp->f_path.dentry->d_name.len,
2163 filp->f_path.dentry->d_name.name);
2172 * buffered reads protect themselves in ->readpage(). O_DIRECT reads
2173 * need locks to protect pending reads from racing with truncate.
2175 if (filp->f_flags & O_DIRECT) {
2176 down_read(&inode->i_alloc_sem);
2179 ret = ocfs2_rw_lock(inode, 0);
2185 /* communicate with ocfs2_dio_end_io */
2186 ocfs2_iocb_set_rw_locked(iocb, rw_level);
2190 * We're fine letting folks race truncates and extending
2191 * writes with read across the cluster, just like they can
2192 * locally. Hence no rw_lock during read.
2194 * Take and drop the meta data lock to update inode fields
2195 * like i_size. This allows the checks down below
2196 * generic_file_aio_read() a chance of actually working.
2198 ret = ocfs2_inode_lock_atime(inode, filp->f_vfsmnt, &lock_level);
2203 ocfs2_inode_unlock(inode, lock_level);
2205 ret = generic_file_aio_read(iocb, iov, nr_segs, iocb->ki_pos);
2207 mlog(0, "generic_file_aio_read returned -EINVAL\n");
2209 /* buffered aio wouldn't have proper lock coverage today */
2210 BUG_ON(ret == -EIOCBQUEUED && !(filp->f_flags & O_DIRECT));
2212 /* see ocfs2_file_aio_write */
2213 if (ret == -EIOCBQUEUED || !ocfs2_iocb_is_rw_locked(iocb)) {
2220 up_read(&inode->i_alloc_sem);
2222 ocfs2_rw_unlock(inode, rw_level);
2228 const struct inode_operations ocfs2_file_iops = {
2229 .setattr = ocfs2_setattr,
2230 .getattr = ocfs2_getattr,
2231 .permission = ocfs2_permission,
2232 .fallocate = ocfs2_fallocate,
2233 .fiemap = ocfs2_fiemap,
2236 const struct inode_operations ocfs2_special_file_iops = {
2237 .setattr = ocfs2_setattr,
2238 .getattr = ocfs2_getattr,
2239 .permission = ocfs2_permission,
2243 * Other than ->lock, keep ocfs2_fops and ocfs2_dops in sync with
2244 * ocfs2_fops_no_plocks and ocfs2_dops_no_plocks!
2246 const struct file_operations ocfs2_fops = {
2247 .llseek = generic_file_llseek,
2248 .read = do_sync_read,
2249 .write = do_sync_write,
2251 .fsync = ocfs2_sync_file,
2252 .release = ocfs2_file_release,
2253 .open = ocfs2_file_open,
2254 .aio_read = ocfs2_file_aio_read,
2255 .aio_write = ocfs2_file_aio_write,
2256 .unlocked_ioctl = ocfs2_ioctl,
2257 #ifdef CONFIG_COMPAT
2258 .compat_ioctl = ocfs2_compat_ioctl,
2261 .flock = ocfs2_flock,
2262 .splice_read = ocfs2_file_splice_read,
2263 .splice_write = ocfs2_file_splice_write,
2266 const struct file_operations ocfs2_dops = {
2267 .llseek = generic_file_llseek,
2268 .read = generic_read_dir,
2269 .readdir = ocfs2_readdir,
2270 .fsync = ocfs2_sync_file,
2271 .release = ocfs2_dir_release,
2272 .open = ocfs2_dir_open,
2273 .unlocked_ioctl = ocfs2_ioctl,
2274 #ifdef CONFIG_COMPAT
2275 .compat_ioctl = ocfs2_compat_ioctl,
2278 .flock = ocfs2_flock,
2282 * POSIX-lockless variants of our file_operations.
2284 * These will be used if the underlying cluster stack does not support
2285 * posix file locking, if the user passes the "localflocks" mount
2286 * option, or if we have a local-only fs.
2288 * ocfs2_flock is in here because all stacks handle UNIX file locks,
2289 * so we still want it in the case of no stack support for
2290 * plocks. Internally, it will do the right thing when asked to ignore
2293 const struct file_operations ocfs2_fops_no_plocks = {
2294 .llseek = generic_file_llseek,
2295 .read = do_sync_read,
2296 .write = do_sync_write,
2298 .fsync = ocfs2_sync_file,
2299 .release = ocfs2_file_release,
2300 .open = ocfs2_file_open,
2301 .aio_read = ocfs2_file_aio_read,
2302 .aio_write = ocfs2_file_aio_write,
2303 .unlocked_ioctl = ocfs2_ioctl,
2304 #ifdef CONFIG_COMPAT
2305 .compat_ioctl = ocfs2_compat_ioctl,
2307 .flock = ocfs2_flock,
2308 .splice_read = ocfs2_file_splice_read,
2309 .splice_write = ocfs2_file_splice_write,
2312 const struct file_operations ocfs2_dops_no_plocks = {
2313 .llseek = generic_file_llseek,
2314 .read = generic_read_dir,
2315 .readdir = ocfs2_readdir,
2316 .fsync = ocfs2_sync_file,
2317 .release = ocfs2_dir_release,
2318 .open = ocfs2_dir_open,
2319 .unlocked_ioctl = ocfs2_ioctl,
2320 #ifdef CONFIG_COMPAT
2321 .compat_ioctl = ocfs2_compat_ioctl,
2323 .flock = ocfs2_flock,