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"
58 #include "buffer_head_io.h"
60 static int ocfs2_sync_inode(struct inode *inode)
62 filemap_fdatawrite(inode->i_mapping);
63 return sync_mapping_buffers(inode->i_mapping);
66 static int ocfs2_file_open(struct inode *inode, struct file *file)
69 int mode = file->f_flags;
70 struct ocfs2_inode_info *oi = OCFS2_I(inode);
72 mlog_entry("(0x%p, 0x%p, '%.*s')\n", inode, file,
73 file->f_path.dentry->d_name.len, file->f_path.dentry->d_name.name);
75 spin_lock(&oi->ip_lock);
77 /* Check that the inode hasn't been wiped from disk by another
78 * node. If it hasn't then we're safe as long as we hold the
79 * spin lock until our increment of open count. */
80 if (OCFS2_I(inode)->ip_flags & OCFS2_INODE_DELETED) {
81 spin_unlock(&oi->ip_lock);
88 oi->ip_flags |= OCFS2_INODE_OPEN_DIRECT;
91 spin_unlock(&oi->ip_lock);
98 static int ocfs2_file_release(struct inode *inode, struct file *file)
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,
104 file->f_path.dentry->d_name.name);
106 spin_lock(&oi->ip_lock);
107 if (!--oi->ip_open_count)
108 oi->ip_flags &= ~OCFS2_INODE_OPEN_DIRECT;
109 spin_unlock(&oi->ip_lock);
116 static int ocfs2_sync_file(struct file *file,
117 struct dentry *dentry,
122 struct inode *inode = dentry->d_inode;
123 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
125 mlog_entry("(0x%p, 0x%p, %d, '%.*s')\n", file, dentry, datasync,
126 dentry->d_name.len, dentry->d_name.name);
128 err = ocfs2_sync_inode(dentry->d_inode);
132 journal = osb->journal->j_journal;
133 err = journal_force_commit(journal);
138 return (err < 0) ? -EIO : 0;
141 int ocfs2_should_update_atime(struct inode *inode,
142 struct vfsmount *vfsmnt)
145 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
147 if (ocfs2_is_hard_readonly(osb) || ocfs2_is_soft_readonly(osb))
150 if ((inode->i_flags & S_NOATIME) ||
151 ((inode->i_sb->s_flags & MS_NODIRATIME) && S_ISDIR(inode->i_mode)))
155 * We can be called with no vfsmnt structure - NFSD will
158 * Note that our action here is different than touch_atime() -
159 * if we can't tell whether this is a noatime mount, then we
160 * don't know whether to trust the value of s_atime_quantum.
165 if ((vfsmnt->mnt_flags & MNT_NOATIME) ||
166 ((vfsmnt->mnt_flags & MNT_NODIRATIME) && S_ISDIR(inode->i_mode)))
169 if (vfsmnt->mnt_flags & MNT_RELATIME) {
170 if ((timespec_compare(&inode->i_atime, &inode->i_mtime) <= 0) ||
171 (timespec_compare(&inode->i_atime, &inode->i_ctime) <= 0))
178 if ((now.tv_sec - inode->i_atime.tv_sec <= osb->s_atime_quantum))
184 int ocfs2_update_inode_atime(struct inode *inode,
185 struct buffer_head *bh)
188 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
190 struct ocfs2_dinode *di = (struct ocfs2_dinode *) bh->b_data;
194 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
195 if (handle == NULL) {
201 ret = ocfs2_journal_access(handle, inode, bh,
202 OCFS2_JOURNAL_ACCESS_WRITE);
209 * Don't use ocfs2_mark_inode_dirty() here as we don't always
210 * have i_mutex to guard against concurrent changes to other
213 inode->i_atime = CURRENT_TIME;
214 di->i_atime = cpu_to_le64(inode->i_atime.tv_sec);
215 di->i_atime_nsec = cpu_to_le32(inode->i_atime.tv_nsec);
217 ret = ocfs2_journal_dirty(handle, bh);
222 ocfs2_commit_trans(OCFS2_SB(inode->i_sb), handle);
228 static int ocfs2_set_inode_size(handle_t *handle,
230 struct buffer_head *fe_bh,
236 i_size_write(inode, new_i_size);
237 inode->i_blocks = ocfs2_inode_sector_count(inode);
238 inode->i_ctime = inode->i_mtime = CURRENT_TIME;
240 status = ocfs2_mark_inode_dirty(handle, inode, fe_bh);
251 static int ocfs2_simple_size_update(struct inode *inode,
252 struct buffer_head *di_bh,
256 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
257 handle_t *handle = NULL;
259 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
260 if (handle == NULL) {
266 ret = ocfs2_set_inode_size(handle, inode, di_bh,
271 ocfs2_commit_trans(osb, handle);
276 static int ocfs2_orphan_for_truncate(struct ocfs2_super *osb,
278 struct buffer_head *fe_bh,
283 struct ocfs2_dinode *di;
288 /* TODO: This needs to actually orphan the inode in this
291 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
292 if (IS_ERR(handle)) {
293 status = PTR_ERR(handle);
298 status = ocfs2_journal_access(handle, inode, fe_bh,
299 OCFS2_JOURNAL_ACCESS_WRITE);
306 * Do this before setting i_size.
308 cluster_bytes = ocfs2_align_bytes_to_clusters(inode->i_sb, new_i_size);
309 status = ocfs2_zero_range_for_truncate(inode, handle, new_i_size,
316 i_size_write(inode, new_i_size);
317 inode->i_blocks = ocfs2_align_bytes_to_sectors(new_i_size);
318 inode->i_ctime = inode->i_mtime = CURRENT_TIME;
320 di = (struct ocfs2_dinode *) fe_bh->b_data;
321 di->i_size = cpu_to_le64(new_i_size);
322 di->i_ctime = di->i_mtime = cpu_to_le64(inode->i_ctime.tv_sec);
323 di->i_ctime_nsec = di->i_mtime_nsec = cpu_to_le32(inode->i_ctime.tv_nsec);
325 status = ocfs2_journal_dirty(handle, fe_bh);
330 ocfs2_commit_trans(osb, handle);
337 static int ocfs2_truncate_file(struct inode *inode,
338 struct buffer_head *di_bh,
342 struct ocfs2_dinode *fe = NULL;
343 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
344 struct ocfs2_truncate_context *tc = NULL;
346 mlog_entry("(inode = %llu, new_i_size = %llu\n",
347 (unsigned long long)OCFS2_I(inode)->ip_blkno,
348 (unsigned long long)new_i_size);
350 fe = (struct ocfs2_dinode *) di_bh->b_data;
351 if (!OCFS2_IS_VALID_DINODE(fe)) {
352 OCFS2_RO_ON_INVALID_DINODE(inode->i_sb, fe);
357 mlog_bug_on_msg(le64_to_cpu(fe->i_size) != i_size_read(inode),
358 "Inode %llu, inode i_size = %lld != di "
359 "i_size = %llu, i_flags = 0x%x\n",
360 (unsigned long long)OCFS2_I(inode)->ip_blkno,
362 (unsigned long long)le64_to_cpu(fe->i_size),
363 le32_to_cpu(fe->i_flags));
365 if (new_i_size > le64_to_cpu(fe->i_size)) {
366 mlog(0, "asked to truncate file with size (%llu) to size (%llu)!\n",
367 (unsigned long long)le64_to_cpu(fe->i_size),
368 (unsigned long long)new_i_size);
374 mlog(0, "inode %llu, i_size = %llu, new_i_size = %llu\n",
375 (unsigned long long)le64_to_cpu(fe->i_blkno),
376 (unsigned long long)le64_to_cpu(fe->i_size),
377 (unsigned long long)new_i_size);
379 /* lets handle the simple truncate cases before doing any more
380 * cluster locking. */
381 if (new_i_size == le64_to_cpu(fe->i_size))
384 down_write(&OCFS2_I(inode)->ip_alloc_sem);
386 /* This forces other nodes to sync and drop their pages. Do
387 * this even if we have a truncate without allocation change -
388 * ocfs2 cluster sizes can be much greater than page size, so
389 * we have to truncate them anyway. */
390 status = ocfs2_data_lock(inode, 1);
392 up_write(&OCFS2_I(inode)->ip_alloc_sem);
398 unmap_mapping_range(inode->i_mapping, new_i_size + PAGE_SIZE - 1, 0, 1);
399 truncate_inode_pages(inode->i_mapping, new_i_size);
401 /* alright, we're going to need to do a full blown alloc size
402 * change. Orphan the inode so that recovery can complete the
403 * truncate if necessary. This does the task of marking
405 status = ocfs2_orphan_for_truncate(osb, inode, di_bh, new_i_size);
408 goto bail_unlock_data;
411 status = ocfs2_prepare_truncate(osb, inode, di_bh, &tc);
414 goto bail_unlock_data;
417 status = ocfs2_commit_truncate(osb, inode, di_bh, tc);
420 goto bail_unlock_data;
423 /* TODO: orphan dir cleanup here. */
425 ocfs2_data_unlock(inode, 1);
427 up_write(&OCFS2_I(inode)->ip_alloc_sem);
436 * extend allocation only here.
437 * we'll update all the disk stuff, and oip->alloc_size
439 * expect stuff to be locked, a transaction started and enough data /
440 * metadata reservations in the contexts.
442 * Will return -EAGAIN, and a reason if a restart is needed.
443 * If passed in, *reason will always be set, even in error.
445 int ocfs2_do_extend_allocation(struct ocfs2_super *osb,
450 struct buffer_head *fe_bh,
452 struct ocfs2_alloc_context *data_ac,
453 struct ocfs2_alloc_context *meta_ac,
454 enum ocfs2_alloc_restarted *reason_ret)
458 struct ocfs2_dinode *fe = (struct ocfs2_dinode *) fe_bh->b_data;
459 enum ocfs2_alloc_restarted reason = RESTART_NONE;
460 u32 bit_off, num_bits;
464 BUG_ON(!clusters_to_add);
467 flags = OCFS2_EXT_UNWRITTEN;
469 free_extents = ocfs2_num_free_extents(osb, inode, fe);
470 if (free_extents < 0) {
471 status = free_extents;
476 /* there are two cases which could cause us to EAGAIN in the
477 * we-need-more-metadata case:
478 * 1) we haven't reserved *any*
479 * 2) we are so fragmented, we've needed to add metadata too
481 if (!free_extents && !meta_ac) {
482 mlog(0, "we haven't reserved any metadata!\n");
484 reason = RESTART_META;
486 } else if ((!free_extents)
487 && (ocfs2_alloc_context_bits_left(meta_ac)
488 < ocfs2_extend_meta_needed(fe))) {
489 mlog(0, "filesystem is really fragmented...\n");
491 reason = RESTART_META;
495 status = ocfs2_claim_clusters(osb, handle, data_ac, 1,
496 &bit_off, &num_bits);
498 if (status != -ENOSPC)
503 BUG_ON(num_bits > clusters_to_add);
505 /* reserve our write early -- insert_extent may update the inode */
506 status = ocfs2_journal_access(handle, inode, fe_bh,
507 OCFS2_JOURNAL_ACCESS_WRITE);
513 block = ocfs2_clusters_to_blocks(osb->sb, bit_off);
514 mlog(0, "Allocating %u clusters at block %u for inode %llu\n",
515 num_bits, bit_off, (unsigned long long)OCFS2_I(inode)->ip_blkno);
516 status = ocfs2_insert_extent(osb, handle, inode, fe_bh,
517 *logical_offset, block, num_bits,
524 status = ocfs2_journal_dirty(handle, fe_bh);
530 clusters_to_add -= num_bits;
531 *logical_offset += num_bits;
533 if (clusters_to_add) {
534 mlog(0, "need to alloc once more, clusters = %u, wanted = "
535 "%u\n", fe->i_clusters, clusters_to_add);
537 reason = RESTART_TRANS;
543 *reason_ret = reason;
548 * For a given allocation, determine which allocators will need to be
549 * accessed, and lock them, reserving the appropriate number of bits.
551 * Sparse file systems call this from ocfs2_write_begin_nolock()
552 * and ocfs2_allocate_unwritten_extents().
554 * File systems which don't support holes call this from
555 * ocfs2_extend_allocation().
557 int ocfs2_lock_allocators(struct inode *inode, struct ocfs2_dinode *di,
558 u32 clusters_to_add, u32 extents_to_split,
559 struct ocfs2_alloc_context **data_ac,
560 struct ocfs2_alloc_context **meta_ac)
562 int ret = 0, num_free_extents;
563 unsigned int max_recs_needed = clusters_to_add + 2 * extents_to_split;
564 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
570 BUG_ON(clusters_to_add != 0 && data_ac == NULL);
572 mlog(0, "extend inode %llu, i_size = %lld, di->i_clusters = %u, "
573 "clusters_to_add = %u, extents_to_split = %u\n",
574 (unsigned long long)OCFS2_I(inode)->ip_blkno, i_size_read(inode),
575 le32_to_cpu(di->i_clusters), clusters_to_add, extents_to_split);
577 num_free_extents = ocfs2_num_free_extents(osb, inode, di);
578 if (num_free_extents < 0) {
579 ret = num_free_extents;
585 * Sparse allocation file systems need to be more conservative
586 * with reserving room for expansion - the actual allocation
587 * happens while we've got a journal handle open so re-taking
588 * a cluster lock (because we ran out of room for another
589 * extent) will violate ordering rules.
591 * Most of the time we'll only be seeing this 1 cluster at a time
594 * Always lock for any unwritten extents - we might want to
595 * add blocks during a split.
597 if (!num_free_extents ||
598 (ocfs2_sparse_alloc(osb) && num_free_extents < max_recs_needed)) {
599 ret = ocfs2_reserve_new_metadata(osb, di, meta_ac);
607 if (clusters_to_add == 0)
610 ret = ocfs2_reserve_clusters(osb, clusters_to_add, data_ac);
620 ocfs2_free_alloc_context(*meta_ac);
625 * We cannot have an error and a non null *data_ac.
632 static int __ocfs2_extend_allocation(struct inode *inode, u32 logical_start,
633 u32 clusters_to_add, int mark_unwritten)
636 int restart_func = 0;
639 struct buffer_head *bh = NULL;
640 struct ocfs2_dinode *fe = NULL;
641 handle_t *handle = NULL;
642 struct ocfs2_alloc_context *data_ac = NULL;
643 struct ocfs2_alloc_context *meta_ac = NULL;
644 enum ocfs2_alloc_restarted why;
645 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
647 mlog_entry("(clusters_to_add = %u)\n", clusters_to_add);
650 * This function only exists for file systems which don't
653 BUG_ON(mark_unwritten && !ocfs2_sparse_alloc(osb));
655 status = ocfs2_read_block(osb, OCFS2_I(inode)->ip_blkno, &bh,
656 OCFS2_BH_CACHED, inode);
662 fe = (struct ocfs2_dinode *) bh->b_data;
663 if (!OCFS2_IS_VALID_DINODE(fe)) {
664 OCFS2_RO_ON_INVALID_DINODE(inode->i_sb, fe);
670 BUG_ON(le32_to_cpu(fe->i_clusters) != OCFS2_I(inode)->ip_clusters);
672 status = ocfs2_lock_allocators(inode, fe, clusters_to_add, 0, &data_ac,
679 credits = ocfs2_calc_extend_credits(osb->sb, fe, clusters_to_add);
680 handle = ocfs2_start_trans(osb, credits);
681 if (IS_ERR(handle)) {
682 status = PTR_ERR(handle);
688 restarted_transaction:
689 /* reserve a write to the file entry early on - that we if we
690 * run out of credits in the allocation path, we can still
692 status = ocfs2_journal_access(handle, inode, bh,
693 OCFS2_JOURNAL_ACCESS_WRITE);
699 prev_clusters = OCFS2_I(inode)->ip_clusters;
701 status = ocfs2_do_extend_allocation(osb,
711 if ((status < 0) && (status != -EAGAIN)) {
712 if (status != -ENOSPC)
717 status = ocfs2_journal_dirty(handle, bh);
723 spin_lock(&OCFS2_I(inode)->ip_lock);
724 clusters_to_add -= (OCFS2_I(inode)->ip_clusters - prev_clusters);
725 spin_unlock(&OCFS2_I(inode)->ip_lock);
727 if (why != RESTART_NONE && clusters_to_add) {
728 if (why == RESTART_META) {
729 mlog(0, "restarting function.\n");
732 BUG_ON(why != RESTART_TRANS);
734 mlog(0, "restarting transaction.\n");
735 /* TODO: This can be more intelligent. */
736 credits = ocfs2_calc_extend_credits(osb->sb,
739 status = ocfs2_extend_trans(handle, credits);
741 /* handle still has to be committed at
747 goto restarted_transaction;
751 mlog(0, "fe: i_clusters = %u, i_size=%llu\n",
752 le32_to_cpu(fe->i_clusters),
753 (unsigned long long)le64_to_cpu(fe->i_size));
754 mlog(0, "inode: ip_clusters=%u, i_size=%lld\n",
755 OCFS2_I(inode)->ip_clusters, i_size_read(inode));
759 ocfs2_commit_trans(osb, handle);
763 ocfs2_free_alloc_context(data_ac);
767 ocfs2_free_alloc_context(meta_ac);
770 if ((!status) && restart_func) {
783 static int ocfs2_extend_allocation(struct inode *inode, u32 logical_start,
784 u32 clusters_to_add, int mark_unwritten)
789 * The alloc sem blocks peope in read/write from reading our
790 * allocation until we're done changing it. We depend on
791 * i_mutex to block other extend/truncate calls while we're
794 down_write(&OCFS2_I(inode)->ip_alloc_sem);
795 ret = __ocfs2_extend_allocation(inode, logical_start, clusters_to_add,
797 up_write(&OCFS2_I(inode)->ip_alloc_sem);
802 /* Some parts of this taken from generic_cont_expand, which turned out
803 * to be too fragile to do exactly what we need without us having to
804 * worry about recursive locking in ->prepare_write() and
805 * ->commit_write(). */
806 static int ocfs2_write_zero_page(struct inode *inode,
809 struct address_space *mapping = inode->i_mapping;
813 handle_t *handle = NULL;
816 offset = (size & (PAGE_CACHE_SIZE-1)); /* Within page */
817 /* ugh. in prepare/commit_write, if from==to==start of block, we
818 ** skip the prepare. make sure we never send an offset for the start
821 if ((offset & (inode->i_sb->s_blocksize - 1)) == 0) {
824 index = size >> PAGE_CACHE_SHIFT;
826 page = grab_cache_page(mapping, index);
833 ret = ocfs2_prepare_write_nolock(inode, page, offset, offset);
839 if (ocfs2_should_order_data(inode)) {
840 handle = ocfs2_start_walk_page_trans(inode, page, offset,
842 if (IS_ERR(handle)) {
843 ret = PTR_ERR(handle);
849 /* must not update i_size! */
850 ret = block_commit_write(page, offset, offset);
857 ocfs2_commit_trans(OCFS2_SB(inode->i_sb), handle);
860 page_cache_release(page);
865 static int ocfs2_zero_extend(struct inode *inode,
870 struct super_block *sb = inode->i_sb;
872 start_off = ocfs2_align_bytes_to_blocks(sb, i_size_read(inode));
873 while (start_off < zero_to_size) {
874 ret = ocfs2_write_zero_page(inode, start_off);
880 start_off += sb->s_blocksize;
883 * Very large extends have the potential to lock up
884 * the cpu for extended periods of time.
894 * A tail_to_skip value > 0 indicates that we're being called from
895 * ocfs2_file_aio_write(). This has the following implications:
897 * - we don't want to update i_size
898 * - di_bh will be NULL, which is fine because it's only used in the
899 * case where we want to update i_size.
900 * - ocfs2_zero_extend() will then only be filling the hole created
901 * between i_size and the start of the write.
903 static int ocfs2_extend_file(struct inode *inode,
904 struct buffer_head *di_bh,
909 u32 clusters_to_add = 0;
911 BUG_ON(!tail_to_skip && !di_bh);
913 /* setattr sometimes calls us like this. */
917 if (i_size_read(inode) == new_i_size)
919 BUG_ON(new_i_size < i_size_read(inode));
921 if (ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb))) {
922 BUG_ON(tail_to_skip != 0);
923 goto out_update_size;
926 clusters_to_add = ocfs2_clusters_for_bytes(inode->i_sb, new_i_size) -
927 OCFS2_I(inode)->ip_clusters;
930 * protect the pages that ocfs2_zero_extend is going to be
931 * pulling into the page cache.. we do this before the
932 * metadata extend so that we don't get into the situation
933 * where we've extended the metadata but can't get the data
936 ret = ocfs2_data_lock(inode, 1);
942 if (clusters_to_add) {
943 ret = ocfs2_extend_allocation(inode,
944 OCFS2_I(inode)->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, (u64)new_i_size - tail_to_skip);
965 /* We're being called from ocfs2_setattr() which wants
966 * us to update i_size */
967 ret = ocfs2_simple_size_update(inode, di_bh, new_i_size);
973 if (!ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)))
974 ocfs2_data_unlock(inode, 1);
980 int ocfs2_setattr(struct dentry *dentry, struct iattr *attr)
982 int status = 0, size_change;
983 struct inode *inode = dentry->d_inode;
984 struct super_block *sb = inode->i_sb;
985 struct ocfs2_super *osb = OCFS2_SB(sb);
986 struct buffer_head *bh = NULL;
987 handle_t *handle = NULL;
989 mlog_entry("(0x%p, '%.*s')\n", dentry,
990 dentry->d_name.len, dentry->d_name.name);
992 if (attr->ia_valid & ATTR_MODE)
993 mlog(0, "mode change: %d\n", attr->ia_mode);
994 if (attr->ia_valid & ATTR_UID)
995 mlog(0, "uid change: %d\n", attr->ia_uid);
996 if (attr->ia_valid & ATTR_GID)
997 mlog(0, "gid change: %d\n", attr->ia_gid);
998 if (attr->ia_valid & ATTR_SIZE)
999 mlog(0, "size change...\n");
1000 if (attr->ia_valid & (ATTR_ATIME | ATTR_MTIME | ATTR_CTIME))
1001 mlog(0, "time change...\n");
1003 #define OCFS2_VALID_ATTRS (ATTR_ATIME | ATTR_MTIME | ATTR_CTIME | ATTR_SIZE \
1004 | ATTR_GID | ATTR_UID | ATTR_MODE)
1005 if (!(attr->ia_valid & OCFS2_VALID_ATTRS)) {
1006 mlog(0, "can't handle attrs: 0x%x\n", attr->ia_valid);
1010 status = inode_change_ok(inode, attr);
1014 size_change = S_ISREG(inode->i_mode) && attr->ia_valid & ATTR_SIZE;
1016 status = ocfs2_rw_lock(inode, 1);
1023 status = ocfs2_meta_lock(inode, &bh, 1);
1025 if (status != -ENOENT)
1027 goto bail_unlock_rw;
1030 if (size_change && attr->ia_size != i_size_read(inode)) {
1031 if (attr->ia_size > sb->s_maxbytes) {
1036 if (i_size_read(inode) > attr->ia_size)
1037 status = ocfs2_truncate_file(inode, bh, attr->ia_size);
1039 status = ocfs2_extend_file(inode, bh, attr->ia_size, 0);
1041 if (status != -ENOSPC)
1048 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1049 if (IS_ERR(handle)) {
1050 status = PTR_ERR(handle);
1056 * This will intentionally not wind up calling vmtruncate(),
1057 * since all the work for a size change has been done above.
1058 * Otherwise, we could get into problems with truncate as
1059 * ip_alloc_sem is used there to protect against i_size
1062 status = inode_setattr(inode, attr);
1068 status = ocfs2_mark_inode_dirty(handle, inode, bh);
1073 ocfs2_commit_trans(osb, handle);
1075 ocfs2_meta_unlock(inode, 1);
1078 ocfs2_rw_unlock(inode, 1);
1087 int ocfs2_getattr(struct vfsmount *mnt,
1088 struct dentry *dentry,
1091 struct inode *inode = dentry->d_inode;
1092 struct super_block *sb = dentry->d_inode->i_sb;
1093 struct ocfs2_super *osb = sb->s_fs_info;
1098 err = ocfs2_inode_revalidate(dentry);
1105 generic_fillattr(inode, stat);
1107 /* We set the blksize from the cluster size for performance */
1108 stat->blksize = osb->s_clustersize;
1116 int ocfs2_permission(struct inode *inode, int mask, struct nameidata *nd)
1122 ret = ocfs2_meta_lock(inode, NULL, 0);
1129 ret = generic_permission(inode, mask, NULL);
1131 ocfs2_meta_unlock(inode, 0);
1137 static int __ocfs2_write_remove_suid(struct inode *inode,
1138 struct buffer_head *bh)
1142 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1143 struct ocfs2_dinode *di;
1145 mlog_entry("(Inode %llu, mode 0%o)\n",
1146 (unsigned long long)OCFS2_I(inode)->ip_blkno, inode->i_mode);
1148 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1149 if (handle == NULL) {
1155 ret = ocfs2_journal_access(handle, inode, bh,
1156 OCFS2_JOURNAL_ACCESS_WRITE);
1162 inode->i_mode &= ~S_ISUID;
1163 if ((inode->i_mode & S_ISGID) && (inode->i_mode & S_IXGRP))
1164 inode->i_mode &= ~S_ISGID;
1166 di = (struct ocfs2_dinode *) bh->b_data;
1167 di->i_mode = cpu_to_le16(inode->i_mode);
1169 ret = ocfs2_journal_dirty(handle, bh);
1174 ocfs2_commit_trans(osb, handle);
1181 * Will look for holes and unwritten extents in the range starting at
1182 * pos for count bytes (inclusive).
1184 static int ocfs2_check_range_for_holes(struct inode *inode, loff_t pos,
1188 unsigned int extent_flags;
1189 u32 cpos, clusters, extent_len, phys_cpos;
1190 struct super_block *sb = inode->i_sb;
1192 cpos = pos >> OCFS2_SB(sb)->s_clustersize_bits;
1193 clusters = ocfs2_clusters_for_bytes(sb, pos + count) - cpos;
1196 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos, &extent_len,
1203 if (phys_cpos == 0 || (extent_flags & OCFS2_EXT_UNWRITTEN)) {
1208 if (extent_len > clusters)
1209 extent_len = clusters;
1211 clusters -= extent_len;
1218 static int ocfs2_write_remove_suid(struct inode *inode)
1221 struct buffer_head *bh = NULL;
1222 struct ocfs2_inode_info *oi = OCFS2_I(inode);
1224 ret = ocfs2_read_block(OCFS2_SB(inode->i_sb),
1225 oi->ip_blkno, &bh, OCFS2_BH_CACHED, inode);
1231 ret = __ocfs2_write_remove_suid(inode, bh);
1238 * Allocate enough extents to cover the region starting at byte offset
1239 * start for len bytes. Existing extents are skipped, any extents
1240 * added are marked as "unwritten".
1242 static int ocfs2_allocate_unwritten_extents(struct inode *inode,
1246 u32 cpos, phys_cpos, clusters, alloc_size;
1249 * We consider both start and len to be inclusive.
1251 cpos = start >> OCFS2_SB(inode->i_sb)->s_clustersize_bits;
1252 clusters = ocfs2_clusters_for_bytes(inode->i_sb, start + len);
1256 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos,
1264 * Hole or existing extent len can be arbitrary, so
1265 * cap it to our own allocation request.
1267 if (alloc_size > clusters)
1268 alloc_size = clusters;
1272 * We already have an allocation at this
1273 * region so we can safely skip it.
1278 ret = __ocfs2_extend_allocation(inode, cpos, alloc_size, 1);
1287 clusters -= alloc_size;
1295 static int __ocfs2_remove_inode_range(struct inode *inode,
1296 struct buffer_head *di_bh,
1297 u32 cpos, u32 phys_cpos, u32 len,
1298 struct ocfs2_cached_dealloc_ctxt *dealloc)
1301 u64 phys_blkno = ocfs2_clusters_to_blocks(inode->i_sb, phys_cpos);
1302 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1303 struct inode *tl_inode = osb->osb_tl_inode;
1305 struct ocfs2_alloc_context *meta_ac = NULL;
1306 struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
1308 ret = ocfs2_lock_allocators(inode, di, 0, 1, NULL, &meta_ac);
1314 mutex_lock(&tl_inode->i_mutex);
1316 if (ocfs2_truncate_log_needs_flush(osb)) {
1317 ret = __ocfs2_flush_truncate_log(osb);
1324 handle = ocfs2_start_trans(osb, OCFS2_REMOVE_EXTENT_CREDITS);
1325 if (handle == NULL) {
1331 ret = ocfs2_journal_access(handle, inode, di_bh,
1332 OCFS2_JOURNAL_ACCESS_WRITE);
1338 ret = ocfs2_remove_extent(inode, di_bh, cpos, len, handle, meta_ac,
1345 OCFS2_I(inode)->ip_clusters -= len;
1346 di->i_clusters = cpu_to_le32(OCFS2_I(inode)->ip_clusters);
1348 ret = ocfs2_journal_dirty(handle, di_bh);
1354 ret = ocfs2_truncate_log_append(osb, handle, phys_blkno, len);
1359 ocfs2_commit_trans(osb, handle);
1361 mutex_unlock(&tl_inode->i_mutex);
1364 ocfs2_free_alloc_context(meta_ac);
1370 * Truncate a byte range, avoiding pages within partial clusters. This
1371 * preserves those pages for the zeroing code to write to.
1373 static void ocfs2_truncate_cluster_pages(struct inode *inode, u64 byte_start,
1376 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1378 struct address_space *mapping = inode->i_mapping;
1380 start = (loff_t)ocfs2_align_bytes_to_clusters(inode->i_sb, byte_start);
1381 end = byte_start + byte_len;
1382 end = end & ~(osb->s_clustersize - 1);
1385 unmap_mapping_range(mapping, start, end - start, 0);
1386 truncate_inode_pages_range(mapping, start, end - 1);
1390 static int ocfs2_zero_partial_clusters(struct inode *inode,
1394 u64 tmpend, end = start + len;
1395 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1396 unsigned int csize = osb->s_clustersize;
1400 * The "start" and "end" values are NOT necessarily part of
1401 * the range whose allocation is being deleted. Rather, this
1402 * is what the user passed in with the request. We must zero
1403 * partial clusters here. There's no need to worry about
1404 * physical allocation - the zeroing code knows to skip holes.
1406 mlog(0, "byte start: %llu, end: %llu\n",
1407 (unsigned long long)start, (unsigned long long)end);
1410 * If both edges are on a cluster boundary then there's no
1411 * zeroing required as the region is part of the allocation to
1414 if ((start & (csize - 1)) == 0 && (end & (csize - 1)) == 0)
1417 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1418 if (handle == NULL) {
1425 * We want to get the byte offset of the end of the 1st cluster.
1427 tmpend = (u64)osb->s_clustersize + (start & ~(osb->s_clustersize - 1));
1431 mlog(0, "1st range: start: %llu, tmpend: %llu\n",
1432 (unsigned long long)start, (unsigned long long)tmpend);
1434 ret = ocfs2_zero_range_for_truncate(inode, handle, start, tmpend);
1440 * This may make start and end equal, but the zeroing
1441 * code will skip any work in that case so there's no
1442 * need to catch it up here.
1444 start = end & ~(osb->s_clustersize - 1);
1446 mlog(0, "2nd range: start: %llu, end: %llu\n",
1447 (unsigned long long)start, (unsigned long long)end);
1449 ret = ocfs2_zero_range_for_truncate(inode, handle, start, end);
1454 ocfs2_commit_trans(osb, handle);
1459 static int ocfs2_remove_inode_range(struct inode *inode,
1460 struct buffer_head *di_bh, u64 byte_start,
1464 u32 trunc_start, trunc_len, cpos, phys_cpos, alloc_size;
1465 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1466 struct ocfs2_cached_dealloc_ctxt dealloc;
1468 ocfs2_init_dealloc_ctxt(&dealloc);
1473 trunc_start = ocfs2_clusters_for_bytes(osb->sb, byte_start);
1474 trunc_len = (byte_start + byte_len) >> osb->s_clustersize_bits;
1475 if (trunc_len >= trunc_start)
1476 trunc_len -= trunc_start;
1480 mlog(0, "Inode: %llu, start: %llu, len: %llu, cstart: %u, clen: %u\n",
1481 (unsigned long long)OCFS2_I(inode)->ip_blkno,
1482 (unsigned long long)byte_start,
1483 (unsigned long long)byte_len, trunc_start, trunc_len);
1485 ret = ocfs2_zero_partial_clusters(inode, byte_start, byte_len);
1493 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos,
1500 if (alloc_size > trunc_len)
1501 alloc_size = trunc_len;
1503 /* Only do work for non-holes */
1504 if (phys_cpos != 0) {
1505 ret = __ocfs2_remove_inode_range(inode, di_bh, cpos,
1506 phys_cpos, alloc_size,
1515 trunc_len -= alloc_size;
1518 ocfs2_truncate_cluster_pages(inode, byte_start, byte_len);
1521 ocfs2_schedule_truncate_log_flush(osb, 1);
1522 ocfs2_run_deallocs(osb, &dealloc);
1528 * Parts of this function taken from xfs_change_file_space()
1530 static int __ocfs2_change_file_space(struct file *file, struct inode *inode,
1531 loff_t f_pos, unsigned int cmd,
1532 struct ocfs2_space_resv *sr,
1538 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1539 struct buffer_head *di_bh = NULL;
1541 unsigned long long max_off = inode->i_sb->s_maxbytes;
1543 if (ocfs2_is_hard_readonly(osb) || ocfs2_is_soft_readonly(osb))
1546 mutex_lock(&inode->i_mutex);
1549 * This prevents concurrent writes on other nodes
1551 ret = ocfs2_rw_lock(inode, 1);
1557 ret = ocfs2_meta_lock(inode, &di_bh, 1);
1563 if (inode->i_flags & (S_IMMUTABLE|S_APPEND)) {
1565 goto out_meta_unlock;
1568 switch (sr->l_whence) {
1569 case 0: /*SEEK_SET*/
1571 case 1: /*SEEK_CUR*/
1572 sr->l_start += f_pos;
1574 case 2: /*SEEK_END*/
1575 sr->l_start += i_size_read(inode);
1579 goto out_meta_unlock;
1583 llen = sr->l_len > 0 ? sr->l_len - 1 : sr->l_len;
1586 || sr->l_start > max_off
1587 || (sr->l_start + llen) < 0
1588 || (sr->l_start + llen) > max_off) {
1590 goto out_meta_unlock;
1592 size = sr->l_start + sr->l_len;
1594 if (cmd == OCFS2_IOC_RESVSP || cmd == OCFS2_IOC_RESVSP64) {
1595 if (sr->l_len <= 0) {
1597 goto out_meta_unlock;
1601 if (file && should_remove_suid(file->f_path.dentry)) {
1602 ret = __ocfs2_write_remove_suid(inode, di_bh);
1605 goto out_meta_unlock;
1609 down_write(&OCFS2_I(inode)->ip_alloc_sem);
1611 case OCFS2_IOC_RESVSP:
1612 case OCFS2_IOC_RESVSP64:
1614 * This takes unsigned offsets, but the signed ones we
1615 * pass have been checked against overflow above.
1617 ret = ocfs2_allocate_unwritten_extents(inode, sr->l_start,
1620 case OCFS2_IOC_UNRESVSP:
1621 case OCFS2_IOC_UNRESVSP64:
1622 ret = ocfs2_remove_inode_range(inode, di_bh, sr->l_start,
1628 up_write(&OCFS2_I(inode)->ip_alloc_sem);
1631 goto out_meta_unlock;
1635 * We update c/mtime for these changes
1637 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1638 if (IS_ERR(handle)) {
1639 ret = PTR_ERR(handle);
1641 goto out_meta_unlock;
1644 if (change_size && i_size_read(inode) < size)
1645 i_size_write(inode, size);
1647 inode->i_ctime = inode->i_mtime = CURRENT_TIME;
1648 ret = ocfs2_mark_inode_dirty(handle, inode, di_bh);
1652 ocfs2_commit_trans(osb, handle);
1656 ocfs2_meta_unlock(inode, 1);
1658 ocfs2_rw_unlock(inode, 1);
1660 mutex_unlock(&inode->i_mutex);
1665 int ocfs2_change_file_space(struct file *file, unsigned int cmd,
1666 struct ocfs2_space_resv *sr)
1668 struct inode *inode = file->f_path.dentry->d_inode;
1669 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);;
1671 if ((cmd == OCFS2_IOC_RESVSP || cmd == OCFS2_IOC_RESVSP64) &&
1672 !ocfs2_writes_unwritten_extents(osb))
1674 else if ((cmd == OCFS2_IOC_UNRESVSP || cmd == OCFS2_IOC_UNRESVSP64) &&
1675 !ocfs2_sparse_alloc(osb))
1678 if (!S_ISREG(inode->i_mode))
1681 if (!(file->f_mode & FMODE_WRITE))
1684 return __ocfs2_change_file_space(file, inode, file->f_pos, cmd, sr, 0);
1687 static long ocfs2_fallocate(struct inode *inode, int mode, loff_t offset,
1690 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1691 struct ocfs2_space_resv sr;
1692 int change_size = 1;
1694 if (!ocfs2_writes_unwritten_extents(osb))
1697 if (S_ISDIR(inode->i_mode))
1700 if (mode & FALLOC_FL_KEEP_SIZE)
1704 sr.l_start = (s64)offset;
1705 sr.l_len = (s64)len;
1707 return __ocfs2_change_file_space(NULL, inode, offset,
1708 OCFS2_IOC_RESVSP64, &sr, change_size);
1711 static int ocfs2_prepare_inode_for_write(struct dentry *dentry,
1717 int ret = 0, meta_level = appending;
1718 struct inode *inode = dentry->d_inode;
1720 loff_t newsize, saved_pos;
1723 * We sample i_size under a read level meta lock to see if our write
1724 * is extending the file, if it is we back off and get a write level
1728 ret = ocfs2_meta_lock(inode, NULL, meta_level);
1735 /* Clear suid / sgid if necessary. We do this here
1736 * instead of later in the write path because
1737 * remove_suid() calls ->setattr without any hint that
1738 * we may have already done our cluster locking. Since
1739 * ocfs2_setattr() *must* take cluster locks to
1740 * proceeed, this will lead us to recursively lock the
1741 * inode. There's also the dinode i_size state which
1742 * can be lost via setattr during extending writes (we
1743 * set inode->i_size at the end of a write. */
1744 if (should_remove_suid(dentry)) {
1745 if (meta_level == 0) {
1746 ocfs2_meta_unlock(inode, meta_level);
1751 ret = ocfs2_write_remove_suid(inode);
1758 /* work on a copy of ppos until we're sure that we won't have
1759 * to recalculate it due to relocking. */
1761 saved_pos = i_size_read(inode);
1762 mlog(0, "O_APPEND: inode->i_size=%llu\n", saved_pos);
1767 if (ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb))) {
1768 loff_t end = saved_pos + count;
1771 * Skip the O_DIRECT checks if we don't need
1774 if (!direct_io || !(*direct_io))
1778 * Allowing concurrent direct writes means
1779 * i_size changes wouldn't be synchronized, so
1780 * one node could wind up truncating another
1783 if (end > i_size_read(inode)) {
1789 * We don't fill holes during direct io, so
1790 * check for them here. If any are found, the
1791 * caller will have to retake some cluster
1792 * locks and initiate the io as buffered.
1794 ret = ocfs2_check_range_for_holes(inode, saved_pos,
1805 * The rest of this loop is concerned with legacy file
1806 * systems which don't support sparse files.
1809 newsize = count + saved_pos;
1811 mlog(0, "pos=%lld newsize=%lld cursize=%lld\n",
1812 (long long) saved_pos, (long long) newsize,
1813 (long long) i_size_read(inode));
1815 /* No need for a higher level metadata lock if we're
1816 * never going past i_size. */
1817 if (newsize <= i_size_read(inode))
1820 if (meta_level == 0) {
1821 ocfs2_meta_unlock(inode, meta_level);
1826 spin_lock(&OCFS2_I(inode)->ip_lock);
1827 clusters = ocfs2_clusters_for_bytes(inode->i_sb, newsize) -
1828 OCFS2_I(inode)->ip_clusters;
1829 spin_unlock(&OCFS2_I(inode)->ip_lock);
1831 mlog(0, "Writing at EOF, may need more allocation: "
1832 "i_size = %lld, newsize = %lld, need %u clusters\n",
1833 (long long) i_size_read(inode), (long long) newsize,
1836 /* We only want to continue the rest of this loop if
1837 * our extend will actually require more
1842 ret = ocfs2_extend_file(inode, NULL, newsize, count);
1855 ocfs2_meta_unlock(inode, meta_level);
1862 ocfs2_set_next_iovec(const struct iovec **iovp, size_t *basep, size_t bytes)
1864 const struct iovec *iov = *iovp;
1865 size_t base = *basep;
1868 int copy = min(bytes, iov->iov_len - base);
1872 if (iov->iov_len == base) {
1881 static struct page * ocfs2_get_write_source(char **ret_src_buf,
1882 const struct iovec *cur_iov,
1886 char *buf = cur_iov->iov_base + iov_offset;
1887 struct page *src_page = NULL;
1890 off = (unsigned long)(buf) & ~PAGE_CACHE_MASK;
1892 if (!segment_eq(get_fs(), KERNEL_DS)) {
1894 * Pull in the user page. We want to do this outside
1895 * of the meta data locks in order to preserve locking
1896 * order in case of page fault.
1898 ret = get_user_pages(current, current->mm,
1899 (unsigned long)buf & PAGE_CACHE_MASK, 1,
1900 0, 0, &src_page, NULL);
1902 *ret_src_buf = kmap(src_page) + off;
1904 src_page = ERR_PTR(-EFAULT);
1912 static void ocfs2_put_write_source(struct page *page)
1916 page_cache_release(page);
1920 static ssize_t ocfs2_file_buffered_write(struct file *file, loff_t *ppos,
1921 const struct iovec *iov,
1922 unsigned long nr_segs,
1924 ssize_t o_direct_written)
1927 ssize_t copied, total = 0;
1928 size_t iov_offset = 0, bytes;
1930 const struct iovec *cur_iov = iov;
1931 struct page *user_page, *page;
1932 char * uninitialized_var(buf);
1937 * handle partial DIO write. Adjust cur_iov if needed.
1939 ocfs2_set_next_iovec(&cur_iov, &iov_offset, o_direct_written);
1944 user_page = ocfs2_get_write_source(&buf, cur_iov, iov_offset);
1945 if (IS_ERR(user_page)) {
1946 ret = PTR_ERR(user_page);
1950 /* Stay within our page boundaries */
1951 bytes = min((PAGE_CACHE_SIZE - ((unsigned long)pos & ~PAGE_CACHE_MASK)),
1952 (PAGE_CACHE_SIZE - ((unsigned long)buf & ~PAGE_CACHE_MASK)));
1953 /* Stay within the vector boundary */
1954 bytes = min_t(size_t, bytes, cur_iov->iov_len - iov_offset);
1955 /* Stay within count */
1956 bytes = min(bytes, count);
1959 ret = ocfs2_write_begin(file, file->f_mapping, pos, bytes, 0,
1966 dst = kmap_atomic(page, KM_USER0);
1967 memcpy(dst + (pos & (loff_t)(PAGE_CACHE_SIZE - 1)), buf, bytes);
1968 kunmap_atomic(dst, KM_USER0);
1969 flush_dcache_page(page);
1970 ocfs2_put_write_source(user_page);
1972 copied = ocfs2_write_end(file, file->f_mapping, pos, bytes,
1973 bytes, page, fsdata);
1981 *ppos = pos + copied;
1984 ocfs2_set_next_iovec(&cur_iov, &iov_offset, copied);
1988 return total ? total : ret;
1991 static ssize_t ocfs2_file_aio_write(struct kiocb *iocb,
1992 const struct iovec *iov,
1993 unsigned long nr_segs,
1996 int ret, direct_io, appending, rw_level, have_alloc_sem = 0;
1997 int can_do_direct, sync = 0;
1998 ssize_t written = 0;
1999 size_t ocount; /* original count */
2000 size_t count; /* after file limit checks */
2001 loff_t *ppos = &iocb->ki_pos;
2002 struct file *file = iocb->ki_filp;
2003 struct inode *inode = file->f_path.dentry->d_inode;
2005 mlog_entry("(0x%p, %u, '%.*s')\n", file,
2006 (unsigned int)nr_segs,
2007 file->f_path.dentry->d_name.len,
2008 file->f_path.dentry->d_name.name);
2010 if (iocb->ki_left == 0)
2013 ret = generic_segment_checks(iov, &nr_segs, &ocount, VERIFY_READ);
2019 vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
2021 appending = file->f_flags & O_APPEND ? 1 : 0;
2022 direct_io = file->f_flags & O_DIRECT ? 1 : 0;
2024 mutex_lock(&inode->i_mutex);
2027 /* to match setattr's i_mutex -> i_alloc_sem -> rw_lock ordering */
2029 down_read(&inode->i_alloc_sem);
2033 /* concurrent O_DIRECT writes are allowed */
2034 rw_level = !direct_io;
2035 ret = ocfs2_rw_lock(inode, rw_level);
2041 can_do_direct = direct_io;
2042 ret = ocfs2_prepare_inode_for_write(file->f_path.dentry, ppos,
2043 iocb->ki_left, appending,
2051 * We can't complete the direct I/O as requested, fall back to
2054 if (direct_io && !can_do_direct) {
2055 ocfs2_rw_unlock(inode, rw_level);
2056 up_read(&inode->i_alloc_sem);
2066 if (!sync && ((file->f_flags & O_SYNC) || IS_SYNC(inode)))
2070 * XXX: Is it ok to execute these checks a second time?
2072 ret = generic_write_checks(file, ppos, &count, S_ISBLK(inode->i_mode));
2077 * Set pos so that sync_page_range_nolock() below understands
2078 * where to start from. We might've moved it around via the
2079 * calls above. The range we want to actually sync starts from
2085 /* communicate with ocfs2_dio_end_io */
2086 ocfs2_iocb_set_rw_locked(iocb, rw_level);
2089 written = generic_file_direct_write(iocb, iov, &nr_segs, *ppos,
2090 ppos, count, ocount);
2096 written = ocfs2_file_buffered_write(file, ppos, iov, nr_segs,
2100 if (ret != -EFAULT || ret != -ENOSPC)
2107 /* buffered aio wouldn't have proper lock coverage today */
2108 BUG_ON(ret == -EIOCBQUEUED && !(file->f_flags & O_DIRECT));
2111 * deep in g_f_a_w_n()->ocfs2_direct_IO we pass in a ocfs2_dio_end_io
2112 * function pointer which is called when o_direct io completes so that
2113 * it can unlock our rw lock. (it's the clustered equivalent of
2114 * i_alloc_sem; protects truncate from racing with pending ios).
2115 * Unfortunately there are error cases which call end_io and others
2116 * that don't. so we don't have to unlock the rw_lock if either an
2117 * async dio is going to do it in the future or an end_io after an
2118 * error has already done it.
2120 if (ret == -EIOCBQUEUED || !ocfs2_iocb_is_rw_locked(iocb)) {
2127 ocfs2_rw_unlock(inode, rw_level);
2131 up_read(&inode->i_alloc_sem);
2133 if (written > 0 && sync) {
2136 err = sync_page_range_nolock(inode, file->f_mapping, pos, count);
2141 mutex_unlock(&inode->i_mutex);
2144 return written ? written : ret;
2147 static int ocfs2_splice_write_actor(struct pipe_inode_info *pipe,
2148 struct pipe_buffer *buf,
2149 struct splice_desc *sd)
2153 struct file *file = sd->u.file;
2154 unsigned int offset;
2155 struct page *page = NULL;
2159 ret = buf->ops->confirm(pipe, buf);
2163 offset = sd->pos & ~PAGE_CACHE_MASK;
2165 if (count + offset > PAGE_CACHE_SIZE)
2166 count = PAGE_CACHE_SIZE - offset;
2168 ret = ocfs2_write_begin(file, file->f_mapping, sd->pos, count, 0,
2175 src = buf->ops->map(pipe, buf, 1);
2176 dst = kmap_atomic(page, KM_USER1);
2177 memcpy(dst + offset, src + buf->offset, count);
2178 kunmap_atomic(dst, KM_USER1);
2179 buf->ops->unmap(pipe, buf, src);
2181 copied = ocfs2_write_end(file, file->f_mapping, sd->pos, count, count,
2190 return copied ? copied : ret;
2193 static ssize_t __ocfs2_file_splice_write(struct pipe_inode_info *pipe,
2200 struct address_space *mapping = out->f_mapping;
2201 struct inode *inode = mapping->host;
2202 struct splice_desc sd = {
2209 ret = __splice_from_pipe(pipe, &sd, ocfs2_splice_write_actor);
2213 if (unlikely((out->f_flags & O_SYNC) || IS_SYNC(inode))) {
2214 err = generic_osync_inode(inode, mapping,
2215 OSYNC_METADATA|OSYNC_DATA);
2224 static ssize_t ocfs2_file_splice_write(struct pipe_inode_info *pipe,
2231 struct inode *inode = out->f_path.dentry->d_inode;
2233 mlog_entry("(0x%p, 0x%p, %u, '%.*s')\n", out, pipe,
2235 out->f_path.dentry->d_name.len,
2236 out->f_path.dentry->d_name.name);
2238 inode_double_lock(inode, pipe->inode);
2240 ret = ocfs2_rw_lock(inode, 1);
2246 ret = ocfs2_prepare_inode_for_write(out->f_path.dentry, ppos, len, 0,
2253 /* ok, we're done with i_size and alloc work */
2254 ret = __ocfs2_file_splice_write(pipe, out, ppos, len, flags);
2257 ocfs2_rw_unlock(inode, 1);
2259 inode_double_unlock(inode, pipe->inode);
2265 static ssize_t ocfs2_file_splice_read(struct file *in,
2267 struct pipe_inode_info *pipe,
2272 struct inode *inode = in->f_path.dentry->d_inode;
2274 mlog_entry("(0x%p, 0x%p, %u, '%.*s')\n", in, pipe,
2276 in->f_path.dentry->d_name.len,
2277 in->f_path.dentry->d_name.name);
2280 * See the comment in ocfs2_file_aio_read()
2282 ret = ocfs2_meta_lock(inode, NULL, 0);
2287 ocfs2_meta_unlock(inode, 0);
2289 ret = generic_file_splice_read(in, ppos, pipe, len, flags);
2296 static ssize_t ocfs2_file_aio_read(struct kiocb *iocb,
2297 const struct iovec *iov,
2298 unsigned long nr_segs,
2301 int ret = 0, rw_level = -1, have_alloc_sem = 0, lock_level = 0;
2302 struct file *filp = iocb->ki_filp;
2303 struct inode *inode = filp->f_path.dentry->d_inode;
2305 mlog_entry("(0x%p, %u, '%.*s')\n", filp,
2306 (unsigned int)nr_segs,
2307 filp->f_path.dentry->d_name.len,
2308 filp->f_path.dentry->d_name.name);
2317 * buffered reads protect themselves in ->readpage(). O_DIRECT reads
2318 * need locks to protect pending reads from racing with truncate.
2320 if (filp->f_flags & O_DIRECT) {
2321 down_read(&inode->i_alloc_sem);
2324 ret = ocfs2_rw_lock(inode, 0);
2330 /* communicate with ocfs2_dio_end_io */
2331 ocfs2_iocb_set_rw_locked(iocb, rw_level);
2335 * We're fine letting folks race truncates and extending
2336 * writes with read across the cluster, just like they can
2337 * locally. Hence no rw_lock during read.
2339 * Take and drop the meta data lock to update inode fields
2340 * like i_size. This allows the checks down below
2341 * generic_file_aio_read() a chance of actually working.
2343 ret = ocfs2_meta_lock_atime(inode, filp->f_vfsmnt, &lock_level);
2348 ocfs2_meta_unlock(inode, lock_level);
2350 ret = generic_file_aio_read(iocb, iov, nr_segs, iocb->ki_pos);
2352 mlog(ML_ERROR, "generic_file_aio_read returned -EINVAL\n");
2354 /* buffered aio wouldn't have proper lock coverage today */
2355 BUG_ON(ret == -EIOCBQUEUED && !(filp->f_flags & O_DIRECT));
2357 /* see ocfs2_file_aio_write */
2358 if (ret == -EIOCBQUEUED || !ocfs2_iocb_is_rw_locked(iocb)) {
2365 up_read(&inode->i_alloc_sem);
2367 ocfs2_rw_unlock(inode, rw_level);
2373 const struct inode_operations ocfs2_file_iops = {
2374 .setattr = ocfs2_setattr,
2375 .getattr = ocfs2_getattr,
2376 .permission = ocfs2_permission,
2377 .fallocate = ocfs2_fallocate,
2380 const struct inode_operations ocfs2_special_file_iops = {
2381 .setattr = ocfs2_setattr,
2382 .getattr = ocfs2_getattr,
2383 .permission = ocfs2_permission,
2386 const struct file_operations ocfs2_fops = {
2387 .read = do_sync_read,
2388 .write = do_sync_write,
2390 .fsync = ocfs2_sync_file,
2391 .release = ocfs2_file_release,
2392 .open = ocfs2_file_open,
2393 .aio_read = ocfs2_file_aio_read,
2394 .aio_write = ocfs2_file_aio_write,
2395 .ioctl = ocfs2_ioctl,
2396 #ifdef CONFIG_COMPAT
2397 .compat_ioctl = ocfs2_compat_ioctl,
2399 .splice_read = ocfs2_file_splice_read,
2400 .splice_write = ocfs2_file_splice_write,
2403 const struct file_operations ocfs2_dops = {
2404 .read = generic_read_dir,
2405 .readdir = ocfs2_readdir,
2406 .fsync = ocfs2_sync_file,
2407 .ioctl = ocfs2_ioctl,
2408 #ifdef CONFIG_COMPAT
2409 .compat_ioctl = ocfs2_compat_ioctl,