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>
38 #include <linux/quotaops.h>
40 #define MLOG_MASK_PREFIX ML_INODE
41 #include <cluster/masklog.h>
49 #include "extent_map.h"
62 #include "refcounttree.h"
64 #include "buffer_head_io.h"
66 static int ocfs2_sync_inode(struct inode *inode)
68 filemap_fdatawrite(inode->i_mapping);
69 return sync_mapping_buffers(inode->i_mapping);
72 static int ocfs2_init_file_private(struct inode *inode, struct file *file)
74 struct ocfs2_file_private *fp;
76 fp = kzalloc(sizeof(struct ocfs2_file_private), GFP_KERNEL);
81 mutex_init(&fp->fp_mutex);
82 ocfs2_file_lock_res_init(&fp->fp_flock, fp);
83 file->private_data = fp;
88 static void ocfs2_free_file_private(struct inode *inode, struct file *file)
90 struct ocfs2_file_private *fp = file->private_data;
91 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
94 ocfs2_simple_drop_lockres(osb, &fp->fp_flock);
95 ocfs2_lock_res_free(&fp->fp_flock);
97 file->private_data = NULL;
101 static int ocfs2_file_open(struct inode *inode, struct file *file)
104 int mode = file->f_flags;
105 struct ocfs2_inode_info *oi = OCFS2_I(inode);
107 mlog_entry("(0x%p, 0x%p, '%.*s')\n", inode, file,
108 file->f_path.dentry->d_name.len, file->f_path.dentry->d_name.name);
110 spin_lock(&oi->ip_lock);
112 /* Check that the inode hasn't been wiped from disk by another
113 * node. If it hasn't then we're safe as long as we hold the
114 * spin lock until our increment of open count. */
115 if (OCFS2_I(inode)->ip_flags & OCFS2_INODE_DELETED) {
116 spin_unlock(&oi->ip_lock);
123 oi->ip_flags |= OCFS2_INODE_OPEN_DIRECT;
126 spin_unlock(&oi->ip_lock);
128 status = ocfs2_init_file_private(inode, file);
131 * We want to set open count back if we're failing the
134 spin_lock(&oi->ip_lock);
136 spin_unlock(&oi->ip_lock);
144 static int ocfs2_file_release(struct inode *inode, struct file *file)
146 struct ocfs2_inode_info *oi = OCFS2_I(inode);
148 mlog_entry("(0x%p, 0x%p, '%.*s')\n", inode, file,
149 file->f_path.dentry->d_name.len,
150 file->f_path.dentry->d_name.name);
152 spin_lock(&oi->ip_lock);
153 if (!--oi->ip_open_count)
154 oi->ip_flags &= ~OCFS2_INODE_OPEN_DIRECT;
155 spin_unlock(&oi->ip_lock);
157 ocfs2_free_file_private(inode, file);
164 static int ocfs2_dir_open(struct inode *inode, struct file *file)
166 return ocfs2_init_file_private(inode, file);
169 static int ocfs2_dir_release(struct inode *inode, struct file *file)
171 ocfs2_free_file_private(inode, file);
175 static int ocfs2_sync_file(struct file *file,
176 struct dentry *dentry,
181 struct inode *inode = dentry->d_inode;
182 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
184 mlog_entry("(0x%p, 0x%p, %d, '%.*s')\n", file, dentry, datasync,
185 dentry->d_name.len, dentry->d_name.name);
187 err = ocfs2_sync_inode(dentry->d_inode);
191 if (datasync && !(inode->i_state & I_DIRTY_DATASYNC))
194 journal = osb->journal->j_journal;
195 err = jbd2_journal_force_commit(journal);
200 return (err < 0) ? -EIO : 0;
203 int ocfs2_should_update_atime(struct inode *inode,
204 struct vfsmount *vfsmnt)
207 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
209 if (ocfs2_is_hard_readonly(osb) || ocfs2_is_soft_readonly(osb))
212 if ((inode->i_flags & S_NOATIME) ||
213 ((inode->i_sb->s_flags & MS_NODIRATIME) && S_ISDIR(inode->i_mode)))
217 * We can be called with no vfsmnt structure - NFSD will
220 * Note that our action here is different than touch_atime() -
221 * if we can't tell whether this is a noatime mount, then we
222 * don't know whether to trust the value of s_atime_quantum.
227 if ((vfsmnt->mnt_flags & MNT_NOATIME) ||
228 ((vfsmnt->mnt_flags & MNT_NODIRATIME) && S_ISDIR(inode->i_mode)))
231 if (vfsmnt->mnt_flags & MNT_RELATIME) {
232 if ((timespec_compare(&inode->i_atime, &inode->i_mtime) <= 0) ||
233 (timespec_compare(&inode->i_atime, &inode->i_ctime) <= 0))
240 if ((now.tv_sec - inode->i_atime.tv_sec <= osb->s_atime_quantum))
246 int ocfs2_update_inode_atime(struct inode *inode,
247 struct buffer_head *bh)
250 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
252 struct ocfs2_dinode *di = (struct ocfs2_dinode *) bh->b_data;
256 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
257 if (IS_ERR(handle)) {
258 ret = PTR_ERR(handle);
263 ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), bh,
264 OCFS2_JOURNAL_ACCESS_WRITE);
271 * Don't use ocfs2_mark_inode_dirty() here as we don't always
272 * have i_mutex to guard against concurrent changes to other
275 inode->i_atime = CURRENT_TIME;
276 di->i_atime = cpu_to_le64(inode->i_atime.tv_sec);
277 di->i_atime_nsec = cpu_to_le32(inode->i_atime.tv_nsec);
279 ret = ocfs2_journal_dirty(handle, bh);
284 ocfs2_commit_trans(OCFS2_SB(inode->i_sb), handle);
290 static int ocfs2_set_inode_size(handle_t *handle,
292 struct buffer_head *fe_bh,
298 i_size_write(inode, new_i_size);
299 inode->i_blocks = ocfs2_inode_sector_count(inode);
300 inode->i_ctime = inode->i_mtime = CURRENT_TIME;
302 status = ocfs2_mark_inode_dirty(handle, inode, fe_bh);
313 int ocfs2_simple_size_update(struct inode *inode,
314 struct buffer_head *di_bh,
318 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
319 handle_t *handle = NULL;
321 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
322 if (IS_ERR(handle)) {
323 ret = PTR_ERR(handle);
328 ret = ocfs2_set_inode_size(handle, inode, di_bh,
333 ocfs2_commit_trans(osb, handle);
338 static int ocfs2_cow_file_pos(struct inode *inode,
339 struct buffer_head *fe_bh,
343 u32 phys, cpos = offset >> OCFS2_SB(inode->i_sb)->s_clustersize_bits;
344 unsigned int num_clusters = 0;
345 unsigned int ext_flags = 0;
348 * If the new offset is aligned to the range of the cluster, there is
349 * no space for ocfs2_zero_range_for_truncate to fill, so no need to
352 if ((offset & (OCFS2_SB(inode->i_sb)->s_clustersize - 1)) == 0)
355 status = ocfs2_get_clusters(inode, cpos, &phys,
356 &num_clusters, &ext_flags);
362 if (!(ext_flags & OCFS2_EXT_REFCOUNTED))
365 return ocfs2_refcount_cow(inode, fe_bh, cpos, 1, cpos+1);
371 static int ocfs2_orphan_for_truncate(struct ocfs2_super *osb,
373 struct buffer_head *fe_bh,
378 struct ocfs2_dinode *di;
384 * We need to CoW the cluster contains the offset if it is reflinked
385 * since we will call ocfs2_zero_range_for_truncate later which will
386 * write "0" from offset to the end of the cluster.
388 status = ocfs2_cow_file_pos(inode, fe_bh, new_i_size);
394 /* TODO: This needs to actually orphan the inode in this
397 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
398 if (IS_ERR(handle)) {
399 status = PTR_ERR(handle);
404 status = ocfs2_journal_access_di(handle, INODE_CACHE(inode), fe_bh,
405 OCFS2_JOURNAL_ACCESS_WRITE);
412 * Do this before setting i_size.
414 cluster_bytes = ocfs2_align_bytes_to_clusters(inode->i_sb, new_i_size);
415 status = ocfs2_zero_range_for_truncate(inode, handle, new_i_size,
422 i_size_write(inode, new_i_size);
423 inode->i_ctime = inode->i_mtime = CURRENT_TIME;
425 di = (struct ocfs2_dinode *) fe_bh->b_data;
426 di->i_size = cpu_to_le64(new_i_size);
427 di->i_ctime = di->i_mtime = cpu_to_le64(inode->i_ctime.tv_sec);
428 di->i_ctime_nsec = di->i_mtime_nsec = cpu_to_le32(inode->i_ctime.tv_nsec);
430 status = ocfs2_journal_dirty(handle, fe_bh);
435 ocfs2_commit_trans(osb, handle);
442 static int ocfs2_truncate_file(struct inode *inode,
443 struct buffer_head *di_bh,
447 struct ocfs2_dinode *fe = NULL;
448 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
449 struct ocfs2_truncate_context *tc = NULL;
451 mlog_entry("(inode = %llu, new_i_size = %llu\n",
452 (unsigned long long)OCFS2_I(inode)->ip_blkno,
453 (unsigned long long)new_i_size);
455 /* We trust di_bh because it comes from ocfs2_inode_lock(), which
456 * already validated it */
457 fe = (struct ocfs2_dinode *) di_bh->b_data;
459 mlog_bug_on_msg(le64_to_cpu(fe->i_size) != i_size_read(inode),
460 "Inode %llu, inode i_size = %lld != di "
461 "i_size = %llu, i_flags = 0x%x\n",
462 (unsigned long long)OCFS2_I(inode)->ip_blkno,
464 (unsigned long long)le64_to_cpu(fe->i_size),
465 le32_to_cpu(fe->i_flags));
467 if (new_i_size > le64_to_cpu(fe->i_size)) {
468 mlog(0, "asked to truncate file with size (%llu) to size (%llu)!\n",
469 (unsigned long long)le64_to_cpu(fe->i_size),
470 (unsigned long long)new_i_size);
476 mlog(0, "inode %llu, i_size = %llu, new_i_size = %llu\n",
477 (unsigned long long)le64_to_cpu(fe->i_blkno),
478 (unsigned long long)le64_to_cpu(fe->i_size),
479 (unsigned long long)new_i_size);
481 /* lets handle the simple truncate cases before doing any more
482 * cluster locking. */
483 if (new_i_size == le64_to_cpu(fe->i_size))
486 down_write(&OCFS2_I(inode)->ip_alloc_sem);
489 * The inode lock forced other nodes to sync and drop their
490 * pages, which (correctly) happens even if we have a truncate
491 * without allocation change - ocfs2 cluster sizes can be much
492 * greater than page size, so we have to truncate them
495 unmap_mapping_range(inode->i_mapping, new_i_size + PAGE_SIZE - 1, 0, 1);
496 truncate_inode_pages(inode->i_mapping, new_i_size);
498 if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
499 status = ocfs2_truncate_inline(inode, di_bh, new_i_size,
500 i_size_read(inode), 1);
504 goto bail_unlock_sem;
507 /* alright, we're going to need to do a full blown alloc size
508 * change. Orphan the inode so that recovery can complete the
509 * truncate if necessary. This does the task of marking
511 status = ocfs2_orphan_for_truncate(osb, inode, di_bh, new_i_size);
514 goto bail_unlock_sem;
517 status = ocfs2_prepare_truncate(osb, inode, di_bh, &tc);
520 goto bail_unlock_sem;
523 status = ocfs2_commit_truncate(osb, inode, di_bh, tc);
526 goto bail_unlock_sem;
529 /* TODO: orphan dir cleanup here. */
531 up_write(&OCFS2_I(inode)->ip_alloc_sem);
540 * extend file allocation only here.
541 * we'll update all the disk stuff, and oip->alloc_size
543 * expect stuff to be locked, a transaction started and enough data /
544 * metadata reservations in the contexts.
546 * Will return -EAGAIN, and a reason if a restart is needed.
547 * If passed in, *reason will always be set, even in error.
549 int ocfs2_add_inode_data(struct ocfs2_super *osb,
554 struct buffer_head *fe_bh,
556 struct ocfs2_alloc_context *data_ac,
557 struct ocfs2_alloc_context *meta_ac,
558 enum ocfs2_alloc_restarted *reason_ret)
561 struct ocfs2_extent_tree et;
563 ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), fe_bh);
564 ret = ocfs2_add_clusters_in_btree(handle, &et, logical_offset,
565 clusters_to_add, mark_unwritten,
566 data_ac, meta_ac, reason_ret);
571 static int __ocfs2_extend_allocation(struct inode *inode, u32 logical_start,
572 u32 clusters_to_add, int mark_unwritten)
575 int restart_func = 0;
578 struct buffer_head *bh = NULL;
579 struct ocfs2_dinode *fe = NULL;
580 handle_t *handle = NULL;
581 struct ocfs2_alloc_context *data_ac = NULL;
582 struct ocfs2_alloc_context *meta_ac = NULL;
583 enum ocfs2_alloc_restarted why;
584 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
585 struct ocfs2_extent_tree et;
588 mlog_entry("(clusters_to_add = %u)\n", clusters_to_add);
591 * This function only exists for file systems which don't
594 BUG_ON(mark_unwritten && !ocfs2_sparse_alloc(osb));
596 status = ocfs2_read_inode_block(inode, &bh);
601 fe = (struct ocfs2_dinode *) bh->b_data;
604 BUG_ON(le32_to_cpu(fe->i_clusters) != OCFS2_I(inode)->ip_clusters);
606 mlog(0, "extend inode %llu, i_size = %lld, di->i_clusters = %u, "
607 "clusters_to_add = %u\n",
608 (unsigned long long)OCFS2_I(inode)->ip_blkno,
609 (long long)i_size_read(inode), le32_to_cpu(fe->i_clusters),
611 ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), bh);
612 status = ocfs2_lock_allocators(inode, &et, clusters_to_add, 0,
619 credits = ocfs2_calc_extend_credits(osb->sb, &fe->id2.i_list,
621 handle = ocfs2_start_trans(osb, credits);
622 if (IS_ERR(handle)) {
623 status = PTR_ERR(handle);
629 restarted_transaction:
630 if (vfs_dq_alloc_space_nodirty(inode, ocfs2_clusters_to_bytes(osb->sb,
637 /* reserve a write to the file entry early on - that we if we
638 * run out of credits in the allocation path, we can still
640 status = ocfs2_journal_access_di(handle, INODE_CACHE(inode), bh,
641 OCFS2_JOURNAL_ACCESS_WRITE);
647 prev_clusters = OCFS2_I(inode)->ip_clusters;
649 status = ocfs2_add_inode_data(osb,
659 if ((status < 0) && (status != -EAGAIN)) {
660 if (status != -ENOSPC)
665 status = ocfs2_journal_dirty(handle, bh);
671 spin_lock(&OCFS2_I(inode)->ip_lock);
672 clusters_to_add -= (OCFS2_I(inode)->ip_clusters - prev_clusters);
673 spin_unlock(&OCFS2_I(inode)->ip_lock);
674 /* Release unused quota reservation */
675 vfs_dq_free_space(inode,
676 ocfs2_clusters_to_bytes(osb->sb, clusters_to_add));
679 if (why != RESTART_NONE && clusters_to_add) {
680 if (why == RESTART_META) {
681 mlog(0, "restarting function.\n");
684 BUG_ON(why != RESTART_TRANS);
686 mlog(0, "restarting transaction.\n");
687 /* TODO: This can be more intelligent. */
688 credits = ocfs2_calc_extend_credits(osb->sb,
691 status = ocfs2_extend_trans(handle, credits);
693 /* handle still has to be committed at
699 goto restarted_transaction;
703 mlog(0, "fe: i_clusters = %u, i_size=%llu\n",
704 le32_to_cpu(fe->i_clusters),
705 (unsigned long long)le64_to_cpu(fe->i_size));
706 mlog(0, "inode: ip_clusters=%u, i_size=%lld\n",
707 OCFS2_I(inode)->ip_clusters, (long long)i_size_read(inode));
710 if (status < 0 && did_quota)
711 vfs_dq_free_space(inode,
712 ocfs2_clusters_to_bytes(osb->sb, clusters_to_add));
714 ocfs2_commit_trans(osb, handle);
718 ocfs2_free_alloc_context(data_ac);
722 ocfs2_free_alloc_context(meta_ac);
725 if ((!status) && restart_func) {
736 /* Some parts of this taken from generic_cont_expand, which turned out
737 * to be too fragile to do exactly what we need without us having to
738 * worry about recursive locking in ->write_begin() and ->write_end(). */
739 static int ocfs2_write_zero_page(struct inode *inode,
742 struct address_space *mapping = inode->i_mapping;
746 handle_t *handle = NULL;
749 offset = (size & (PAGE_CACHE_SIZE-1)); /* Within page */
750 /* ugh. in prepare/commit_write, if from==to==start of block, we
751 ** skip the prepare. make sure we never send an offset for the start
754 if ((offset & (inode->i_sb->s_blocksize - 1)) == 0) {
757 index = size >> PAGE_CACHE_SHIFT;
759 page = grab_cache_page(mapping, index);
766 ret = ocfs2_prepare_write_nolock(inode, page, offset, offset);
772 if (ocfs2_should_order_data(inode)) {
773 handle = ocfs2_start_walk_page_trans(inode, page, offset,
775 if (IS_ERR(handle)) {
776 ret = PTR_ERR(handle);
782 /* must not update i_size! */
783 ret = block_commit_write(page, offset, offset);
790 ocfs2_commit_trans(OCFS2_SB(inode->i_sb), handle);
793 page_cache_release(page);
798 static int ocfs2_zero_extend(struct inode *inode,
803 struct super_block *sb = inode->i_sb;
805 start_off = ocfs2_align_bytes_to_blocks(sb, i_size_read(inode));
806 while (start_off < zero_to_size) {
807 ret = ocfs2_write_zero_page(inode, start_off);
813 start_off += sb->s_blocksize;
816 * Very large extends have the potential to lock up
817 * the cpu for extended periods of time.
826 int ocfs2_extend_no_holes(struct inode *inode, u64 new_i_size, u64 zero_to)
830 struct ocfs2_inode_info *oi = OCFS2_I(inode);
832 clusters_to_add = ocfs2_clusters_for_bytes(inode->i_sb, new_i_size);
833 if (clusters_to_add < oi->ip_clusters)
836 clusters_to_add -= oi->ip_clusters;
838 if (clusters_to_add) {
839 ret = __ocfs2_extend_allocation(inode, oi->ip_clusters,
848 * Call this even if we don't add any clusters to the tree. We
849 * still need to zero the area between the old i_size and the
852 ret = ocfs2_zero_extend(inode, zero_to);
860 static int ocfs2_extend_file(struct inode *inode,
861 struct buffer_head *di_bh,
865 struct ocfs2_inode_info *oi = OCFS2_I(inode);
869 /* setattr sometimes calls us like this. */
873 if (i_size_read(inode) == new_i_size)
875 BUG_ON(new_i_size < i_size_read(inode));
878 * Fall through for converting inline data, even if the fs
879 * supports sparse files.
881 * The check for inline data here is legal - nobody can add
882 * the feature since we have i_mutex. We must check it again
883 * after acquiring ip_alloc_sem though, as paths like mmap
884 * might have raced us to converting the inode to extents.
886 if (!(oi->ip_dyn_features & OCFS2_INLINE_DATA_FL)
887 && ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)))
888 goto out_update_size;
891 * The alloc sem blocks people in read/write from reading our
892 * allocation until we're done changing it. We depend on
893 * i_mutex to block other extend/truncate calls while we're
896 down_write(&oi->ip_alloc_sem);
898 if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
900 * We can optimize small extends by keeping the inodes
903 if (ocfs2_size_fits_inline_data(di_bh, new_i_size)) {
904 up_write(&oi->ip_alloc_sem);
905 goto out_update_size;
908 ret = ocfs2_convert_inline_data_to_extents(inode, di_bh);
910 up_write(&oi->ip_alloc_sem);
917 if (!ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)))
918 ret = ocfs2_extend_no_holes(inode, new_i_size, new_i_size);
920 up_write(&oi->ip_alloc_sem);
928 ret = ocfs2_simple_size_update(inode, di_bh, new_i_size);
936 int ocfs2_setattr(struct dentry *dentry, struct iattr *attr)
938 int status = 0, size_change;
939 struct inode *inode = dentry->d_inode;
940 struct super_block *sb = inode->i_sb;
941 struct ocfs2_super *osb = OCFS2_SB(sb);
942 struct buffer_head *bh = NULL;
943 handle_t *handle = NULL;
945 struct dquot *transfer_from[MAXQUOTAS] = { };
946 struct dquot *transfer_to[MAXQUOTAS] = { };
948 mlog_entry("(0x%p, '%.*s')\n", dentry,
949 dentry->d_name.len, dentry->d_name.name);
951 /* ensuring we don't even attempt to truncate a symlink */
952 if (S_ISLNK(inode->i_mode))
953 attr->ia_valid &= ~ATTR_SIZE;
955 if (attr->ia_valid & ATTR_MODE)
956 mlog(0, "mode change: %d\n", attr->ia_mode);
957 if (attr->ia_valid & ATTR_UID)
958 mlog(0, "uid change: %d\n", attr->ia_uid);
959 if (attr->ia_valid & ATTR_GID)
960 mlog(0, "gid change: %d\n", attr->ia_gid);
961 if (attr->ia_valid & ATTR_SIZE)
962 mlog(0, "size change...\n");
963 if (attr->ia_valid & (ATTR_ATIME | ATTR_MTIME | ATTR_CTIME))
964 mlog(0, "time change...\n");
966 #define OCFS2_VALID_ATTRS (ATTR_ATIME | ATTR_MTIME | ATTR_CTIME | ATTR_SIZE \
967 | ATTR_GID | ATTR_UID | ATTR_MODE)
968 if (!(attr->ia_valid & OCFS2_VALID_ATTRS)) {
969 mlog(0, "can't handle attrs: 0x%x\n", attr->ia_valid);
973 status = inode_change_ok(inode, attr);
977 size_change = S_ISREG(inode->i_mode) && attr->ia_valid & ATTR_SIZE;
979 status = ocfs2_rw_lock(inode, 1);
986 status = ocfs2_inode_lock(inode, &bh, 1);
988 if (status != -ENOENT)
993 if (size_change && attr->ia_size != i_size_read(inode)) {
994 if (attr->ia_size > sb->s_maxbytes) {
999 if (i_size_read(inode) > attr->ia_size) {
1000 if (ocfs2_should_order_data(inode)) {
1001 status = ocfs2_begin_ordered_truncate(inode,
1006 status = ocfs2_truncate_file(inode, bh, attr->ia_size);
1008 status = ocfs2_extend_file(inode, bh, attr->ia_size);
1010 if (status != -ENOSPC)
1017 if ((attr->ia_valid & ATTR_UID && attr->ia_uid != inode->i_uid) ||
1018 (attr->ia_valid & ATTR_GID && attr->ia_gid != inode->i_gid)) {
1020 * Gather pointers to quota structures so that allocation /
1021 * freeing of quota structures happens here and not inside
1022 * vfs_dq_transfer() where we have problems with lock ordering
1024 if (attr->ia_valid & ATTR_UID && attr->ia_uid != inode->i_uid
1025 && OCFS2_HAS_RO_COMPAT_FEATURE(sb,
1026 OCFS2_FEATURE_RO_COMPAT_USRQUOTA)) {
1027 transfer_to[USRQUOTA] = dqget(sb, attr->ia_uid,
1029 transfer_from[USRQUOTA] = dqget(sb, inode->i_uid,
1031 if (!transfer_to[USRQUOTA] || !transfer_from[USRQUOTA]) {
1036 if (attr->ia_valid & ATTR_GID && attr->ia_gid != inode->i_gid
1037 && OCFS2_HAS_RO_COMPAT_FEATURE(sb,
1038 OCFS2_FEATURE_RO_COMPAT_GRPQUOTA)) {
1039 transfer_to[GRPQUOTA] = dqget(sb, attr->ia_gid,
1041 transfer_from[GRPQUOTA] = dqget(sb, inode->i_gid,
1043 if (!transfer_to[GRPQUOTA] || !transfer_from[GRPQUOTA]) {
1048 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS +
1049 2 * ocfs2_quota_trans_credits(sb));
1050 if (IS_ERR(handle)) {
1051 status = PTR_ERR(handle);
1055 status = vfs_dq_transfer(inode, attr) ? -EDQUOT : 0;
1059 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1060 if (IS_ERR(handle)) {
1061 status = PTR_ERR(handle);
1068 * This will intentionally not wind up calling vmtruncate(),
1069 * since all the work for a size change has been done above.
1070 * Otherwise, we could get into problems with truncate as
1071 * ip_alloc_sem is used there to protect against i_size
1074 status = inode_setattr(inode, attr);
1080 status = ocfs2_mark_inode_dirty(handle, inode, bh);
1085 ocfs2_commit_trans(osb, handle);
1087 ocfs2_inode_unlock(inode, 1);
1090 ocfs2_rw_unlock(inode, 1);
1094 /* Release quota pointers in case we acquired them */
1095 for (qtype = 0; qtype < MAXQUOTAS; qtype++) {
1096 dqput(transfer_to[qtype]);
1097 dqput(transfer_from[qtype]);
1100 if (!status && attr->ia_valid & ATTR_MODE) {
1101 status = ocfs2_acl_chmod(inode);
1110 int ocfs2_getattr(struct vfsmount *mnt,
1111 struct dentry *dentry,
1114 struct inode *inode = dentry->d_inode;
1115 struct super_block *sb = dentry->d_inode->i_sb;
1116 struct ocfs2_super *osb = sb->s_fs_info;
1121 err = ocfs2_inode_revalidate(dentry);
1128 generic_fillattr(inode, stat);
1130 /* We set the blksize from the cluster size for performance */
1131 stat->blksize = osb->s_clustersize;
1139 int ocfs2_permission(struct inode *inode, int mask)
1145 ret = ocfs2_inode_lock(inode, NULL, 0);
1152 ret = generic_permission(inode, mask, ocfs2_check_acl);
1154 ocfs2_inode_unlock(inode, 0);
1160 static int __ocfs2_write_remove_suid(struct inode *inode,
1161 struct buffer_head *bh)
1165 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1166 struct ocfs2_dinode *di;
1168 mlog_entry("(Inode %llu, mode 0%o)\n",
1169 (unsigned long long)OCFS2_I(inode)->ip_blkno, inode->i_mode);
1171 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1172 if (IS_ERR(handle)) {
1173 ret = PTR_ERR(handle);
1178 ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), bh,
1179 OCFS2_JOURNAL_ACCESS_WRITE);
1185 inode->i_mode &= ~S_ISUID;
1186 if ((inode->i_mode & S_ISGID) && (inode->i_mode & S_IXGRP))
1187 inode->i_mode &= ~S_ISGID;
1189 di = (struct ocfs2_dinode *) bh->b_data;
1190 di->i_mode = cpu_to_le16(inode->i_mode);
1192 ret = ocfs2_journal_dirty(handle, bh);
1197 ocfs2_commit_trans(osb, handle);
1204 * Will look for holes and unwritten extents in the range starting at
1205 * pos for count bytes (inclusive).
1207 static int ocfs2_check_range_for_holes(struct inode *inode, loff_t pos,
1211 unsigned int extent_flags;
1212 u32 cpos, clusters, extent_len, phys_cpos;
1213 struct super_block *sb = inode->i_sb;
1215 cpos = pos >> OCFS2_SB(sb)->s_clustersize_bits;
1216 clusters = ocfs2_clusters_for_bytes(sb, pos + count) - cpos;
1219 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos, &extent_len,
1226 if (phys_cpos == 0 || (extent_flags & OCFS2_EXT_UNWRITTEN)) {
1231 if (extent_len > clusters)
1232 extent_len = clusters;
1234 clusters -= extent_len;
1241 static int ocfs2_write_remove_suid(struct inode *inode)
1244 struct buffer_head *bh = NULL;
1246 ret = ocfs2_read_inode_block(inode, &bh);
1252 ret = __ocfs2_write_remove_suid(inode, bh);
1259 * Allocate enough extents to cover the region starting at byte offset
1260 * start for len bytes. Existing extents are skipped, any extents
1261 * added are marked as "unwritten".
1263 static int ocfs2_allocate_unwritten_extents(struct inode *inode,
1267 u32 cpos, phys_cpos, clusters, alloc_size;
1268 u64 end = start + len;
1269 struct buffer_head *di_bh = NULL;
1271 if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1272 ret = ocfs2_read_inode_block(inode, &di_bh);
1279 * Nothing to do if the requested reservation range
1280 * fits within the inode.
1282 if (ocfs2_size_fits_inline_data(di_bh, end))
1285 ret = ocfs2_convert_inline_data_to_extents(inode, di_bh);
1293 * We consider both start and len to be inclusive.
1295 cpos = start >> OCFS2_SB(inode->i_sb)->s_clustersize_bits;
1296 clusters = ocfs2_clusters_for_bytes(inode->i_sb, start + len);
1300 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos,
1308 * Hole or existing extent len can be arbitrary, so
1309 * cap it to our own allocation request.
1311 if (alloc_size > clusters)
1312 alloc_size = clusters;
1316 * We already have an allocation at this
1317 * region so we can safely skip it.
1322 ret = __ocfs2_extend_allocation(inode, cpos, alloc_size, 1);
1331 clusters -= alloc_size;
1342 * Truncate a byte range, avoiding pages within partial clusters. This
1343 * preserves those pages for the zeroing code to write to.
1345 static void ocfs2_truncate_cluster_pages(struct inode *inode, u64 byte_start,
1348 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1350 struct address_space *mapping = inode->i_mapping;
1352 start = (loff_t)ocfs2_align_bytes_to_clusters(inode->i_sb, byte_start);
1353 end = byte_start + byte_len;
1354 end = end & ~(osb->s_clustersize - 1);
1357 unmap_mapping_range(mapping, start, end - start, 0);
1358 truncate_inode_pages_range(mapping, start, end - 1);
1362 static int ocfs2_zero_partial_clusters(struct inode *inode,
1366 u64 tmpend, end = start + len;
1367 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1368 unsigned int csize = osb->s_clustersize;
1372 * The "start" and "end" values are NOT necessarily part of
1373 * the range whose allocation is being deleted. Rather, this
1374 * is what the user passed in with the request. We must zero
1375 * partial clusters here. There's no need to worry about
1376 * physical allocation - the zeroing code knows to skip holes.
1378 mlog(0, "byte start: %llu, end: %llu\n",
1379 (unsigned long long)start, (unsigned long long)end);
1382 * If both edges are on a cluster boundary then there's no
1383 * zeroing required as the region is part of the allocation to
1386 if ((start & (csize - 1)) == 0 && (end & (csize - 1)) == 0)
1389 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1390 if (IS_ERR(handle)) {
1391 ret = PTR_ERR(handle);
1397 * We want to get the byte offset of the end of the 1st cluster.
1399 tmpend = (u64)osb->s_clustersize + (start & ~(osb->s_clustersize - 1));
1403 mlog(0, "1st range: start: %llu, tmpend: %llu\n",
1404 (unsigned long long)start, (unsigned long long)tmpend);
1406 ret = ocfs2_zero_range_for_truncate(inode, handle, start, tmpend);
1412 * This may make start and end equal, but the zeroing
1413 * code will skip any work in that case so there's no
1414 * need to catch it up here.
1416 start = end & ~(osb->s_clustersize - 1);
1418 mlog(0, "2nd range: start: %llu, end: %llu\n",
1419 (unsigned long long)start, (unsigned long long)end);
1421 ret = ocfs2_zero_range_for_truncate(inode, handle, start, end);
1426 ocfs2_commit_trans(osb, handle);
1431 static int ocfs2_remove_inode_range(struct inode *inode,
1432 struct buffer_head *di_bh, u64 byte_start,
1436 u32 trunc_start, trunc_len, cpos, phys_cpos, alloc_size;
1437 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1438 struct ocfs2_cached_dealloc_ctxt dealloc;
1439 struct address_space *mapping = inode->i_mapping;
1440 struct ocfs2_extent_tree et;
1442 ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), di_bh);
1443 ocfs2_init_dealloc_ctxt(&dealloc);
1448 if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1449 ret = ocfs2_truncate_inline(inode, di_bh, byte_start,
1450 byte_start + byte_len, 0);
1456 * There's no need to get fancy with the page cache
1457 * truncate of an inline-data inode. We're talking
1458 * about less than a page here, which will be cached
1459 * in the dinode buffer anyway.
1461 unmap_mapping_range(mapping, 0, 0, 0);
1462 truncate_inode_pages(mapping, 0);
1466 trunc_start = ocfs2_clusters_for_bytes(osb->sb, byte_start);
1467 trunc_len = (byte_start + byte_len) >> osb->s_clustersize_bits;
1468 if (trunc_len >= trunc_start)
1469 trunc_len -= trunc_start;
1473 mlog(0, "Inode: %llu, start: %llu, len: %llu, cstart: %u, clen: %u\n",
1474 (unsigned long long)OCFS2_I(inode)->ip_blkno,
1475 (unsigned long long)byte_start,
1476 (unsigned long long)byte_len, trunc_start, trunc_len);
1478 ret = ocfs2_zero_partial_clusters(inode, byte_start, byte_len);
1486 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos,
1493 if (alloc_size > trunc_len)
1494 alloc_size = trunc_len;
1496 /* Only do work for non-holes */
1497 if (phys_cpos != 0) {
1498 ret = ocfs2_remove_btree_range(inode, &et, cpos,
1499 phys_cpos, alloc_size,
1508 trunc_len -= alloc_size;
1511 ocfs2_truncate_cluster_pages(inode, byte_start, byte_len);
1514 ocfs2_schedule_truncate_log_flush(osb, 1);
1515 ocfs2_run_deallocs(osb, &dealloc);
1521 * Parts of this function taken from xfs_change_file_space()
1523 static int __ocfs2_change_file_space(struct file *file, struct inode *inode,
1524 loff_t f_pos, unsigned int cmd,
1525 struct ocfs2_space_resv *sr,
1531 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1532 struct buffer_head *di_bh = NULL;
1534 unsigned long long max_off = inode->i_sb->s_maxbytes;
1536 if (ocfs2_is_hard_readonly(osb) || ocfs2_is_soft_readonly(osb))
1539 mutex_lock(&inode->i_mutex);
1542 * This prevents concurrent writes on other nodes
1544 ret = ocfs2_rw_lock(inode, 1);
1550 ret = ocfs2_inode_lock(inode, &di_bh, 1);
1556 if (inode->i_flags & (S_IMMUTABLE|S_APPEND)) {
1558 goto out_inode_unlock;
1561 switch (sr->l_whence) {
1562 case 0: /*SEEK_SET*/
1564 case 1: /*SEEK_CUR*/
1565 sr->l_start += f_pos;
1567 case 2: /*SEEK_END*/
1568 sr->l_start += i_size_read(inode);
1572 goto out_inode_unlock;
1576 llen = sr->l_len > 0 ? sr->l_len - 1 : sr->l_len;
1579 || sr->l_start > max_off
1580 || (sr->l_start + llen) < 0
1581 || (sr->l_start + llen) > max_off) {
1583 goto out_inode_unlock;
1585 size = sr->l_start + sr->l_len;
1587 if (cmd == OCFS2_IOC_RESVSP || cmd == OCFS2_IOC_RESVSP64) {
1588 if (sr->l_len <= 0) {
1590 goto out_inode_unlock;
1594 if (file && should_remove_suid(file->f_path.dentry)) {
1595 ret = __ocfs2_write_remove_suid(inode, di_bh);
1598 goto out_inode_unlock;
1602 down_write(&OCFS2_I(inode)->ip_alloc_sem);
1604 case OCFS2_IOC_RESVSP:
1605 case OCFS2_IOC_RESVSP64:
1607 * This takes unsigned offsets, but the signed ones we
1608 * pass have been checked against overflow above.
1610 ret = ocfs2_allocate_unwritten_extents(inode, sr->l_start,
1613 case OCFS2_IOC_UNRESVSP:
1614 case OCFS2_IOC_UNRESVSP64:
1615 ret = ocfs2_remove_inode_range(inode, di_bh, sr->l_start,
1621 up_write(&OCFS2_I(inode)->ip_alloc_sem);
1624 goto out_inode_unlock;
1628 * We update c/mtime for these changes
1630 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1631 if (IS_ERR(handle)) {
1632 ret = PTR_ERR(handle);
1634 goto out_inode_unlock;
1637 if (change_size && i_size_read(inode) < size)
1638 i_size_write(inode, size);
1640 inode->i_ctime = inode->i_mtime = CURRENT_TIME;
1641 ret = ocfs2_mark_inode_dirty(handle, inode, di_bh);
1645 ocfs2_commit_trans(osb, handle);
1649 ocfs2_inode_unlock(inode, 1);
1651 ocfs2_rw_unlock(inode, 1);
1654 mutex_unlock(&inode->i_mutex);
1658 int ocfs2_change_file_space(struct file *file, unsigned int cmd,
1659 struct ocfs2_space_resv *sr)
1661 struct inode *inode = file->f_path.dentry->d_inode;
1662 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1664 if ((cmd == OCFS2_IOC_RESVSP || cmd == OCFS2_IOC_RESVSP64) &&
1665 !ocfs2_writes_unwritten_extents(osb))
1667 else if ((cmd == OCFS2_IOC_UNRESVSP || cmd == OCFS2_IOC_UNRESVSP64) &&
1668 !ocfs2_sparse_alloc(osb))
1671 if (!S_ISREG(inode->i_mode))
1674 if (!(file->f_mode & FMODE_WRITE))
1677 return __ocfs2_change_file_space(file, inode, file->f_pos, cmd, sr, 0);
1680 static long ocfs2_fallocate(struct inode *inode, int mode, loff_t offset,
1683 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1684 struct ocfs2_space_resv sr;
1685 int change_size = 1;
1687 if (!ocfs2_writes_unwritten_extents(osb))
1690 if (S_ISDIR(inode->i_mode))
1693 if (mode & FALLOC_FL_KEEP_SIZE)
1697 sr.l_start = (s64)offset;
1698 sr.l_len = (s64)len;
1700 return __ocfs2_change_file_space(NULL, inode, offset,
1701 OCFS2_IOC_RESVSP64, &sr, change_size);
1704 int ocfs2_check_range_for_refcount(struct inode *inode, loff_t pos,
1708 unsigned int extent_flags;
1709 u32 cpos, clusters, extent_len, phys_cpos;
1710 struct super_block *sb = inode->i_sb;
1712 if (!ocfs2_refcount_tree(OCFS2_SB(inode->i_sb)) ||
1713 !(OCFS2_I(inode)->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL))
1716 cpos = pos >> OCFS2_SB(sb)->s_clustersize_bits;
1717 clusters = ocfs2_clusters_for_bytes(sb, pos + count) - cpos;
1720 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos, &extent_len,
1727 if (phys_cpos && (extent_flags & OCFS2_EXT_REFCOUNTED)) {
1732 if (extent_len > clusters)
1733 extent_len = clusters;
1735 clusters -= extent_len;
1742 static int ocfs2_prepare_inode_for_refcount(struct inode *inode,
1743 loff_t pos, size_t count,
1747 struct buffer_head *di_bh = NULL;
1748 u32 cpos = pos >> OCFS2_SB(inode->i_sb)->s_clustersize_bits;
1750 ocfs2_clusters_for_bytes(inode->i_sb, pos + count) - cpos;
1752 ret = ocfs2_inode_lock(inode, &di_bh, 1);
1760 ret = ocfs2_refcount_cow(inode, di_bh, cpos, clusters, UINT_MAX);
1768 static int ocfs2_prepare_inode_for_write(struct dentry *dentry,
1774 int ret = 0, meta_level = 0;
1775 struct inode *inode = dentry->d_inode;
1776 loff_t saved_pos, end;
1779 * We start with a read level meta lock and only jump to an ex
1780 * if we need to make modifications here.
1783 ret = ocfs2_inode_lock(inode, NULL, meta_level);
1790 /* Clear suid / sgid if necessary. We do this here
1791 * instead of later in the write path because
1792 * remove_suid() calls ->setattr without any hint that
1793 * we may have already done our cluster locking. Since
1794 * ocfs2_setattr() *must* take cluster locks to
1795 * proceeed, this will lead us to recursively lock the
1796 * inode. There's also the dinode i_size state which
1797 * can be lost via setattr during extending writes (we
1798 * set inode->i_size at the end of a write. */
1799 if (should_remove_suid(dentry)) {
1800 if (meta_level == 0) {
1801 ocfs2_inode_unlock(inode, meta_level);
1806 ret = ocfs2_write_remove_suid(inode);
1813 /* work on a copy of ppos until we're sure that we won't have
1814 * to recalculate it due to relocking. */
1816 saved_pos = i_size_read(inode);
1817 mlog(0, "O_APPEND: inode->i_size=%llu\n", saved_pos);
1822 end = saved_pos + count;
1824 ret = ocfs2_check_range_for_refcount(inode, saved_pos, count);
1826 ocfs2_inode_unlock(inode, meta_level);
1829 ret = ocfs2_prepare_inode_for_refcount(inode,
1841 * Skip the O_DIRECT checks if we don't need
1844 if (!direct_io || !(*direct_io))
1848 * There's no sane way to do direct writes to an inode
1851 if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1857 * Allowing concurrent direct writes means
1858 * i_size changes wouldn't be synchronized, so
1859 * one node could wind up truncating another
1862 if (end > i_size_read(inode)) {
1868 * We don't fill holes during direct io, so
1869 * check for them here. If any are found, the
1870 * caller will have to retake some cluster
1871 * locks and initiate the io as buffered.
1873 ret = ocfs2_check_range_for_holes(inode, saved_pos, count);
1886 if (meta_level >= 0)
1887 ocfs2_inode_unlock(inode, meta_level);
1893 static ssize_t ocfs2_file_aio_write(struct kiocb *iocb,
1894 const struct iovec *iov,
1895 unsigned long nr_segs,
1898 int ret, direct_io, appending, rw_level, have_alloc_sem = 0;
1900 ssize_t written = 0;
1901 size_t ocount; /* original count */
1902 size_t count; /* after file limit checks */
1903 loff_t old_size, *ppos = &iocb->ki_pos;
1905 struct file *file = iocb->ki_filp;
1906 struct inode *inode = file->f_path.dentry->d_inode;
1907 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1909 mlog_entry("(0x%p, %u, '%.*s')\n", file,
1910 (unsigned int)nr_segs,
1911 file->f_path.dentry->d_name.len,
1912 file->f_path.dentry->d_name.name);
1914 if (iocb->ki_left == 0)
1917 vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
1919 appending = file->f_flags & O_APPEND ? 1 : 0;
1920 direct_io = file->f_flags & O_DIRECT ? 1 : 0;
1922 mutex_lock(&inode->i_mutex);
1925 /* to match setattr's i_mutex -> i_alloc_sem -> rw_lock ordering */
1927 down_read(&inode->i_alloc_sem);
1931 /* concurrent O_DIRECT writes are allowed */
1932 rw_level = !direct_io;
1933 ret = ocfs2_rw_lock(inode, rw_level);
1939 can_do_direct = direct_io;
1940 ret = ocfs2_prepare_inode_for_write(file->f_path.dentry, ppos,
1941 iocb->ki_left, appending,
1949 * We can't complete the direct I/O as requested, fall back to
1952 if (direct_io && !can_do_direct) {
1953 ocfs2_rw_unlock(inode, rw_level);
1954 up_read(&inode->i_alloc_sem);
1964 * To later detect whether a journal commit for sync writes is
1965 * necessary, we sample i_size, and cluster count here.
1967 old_size = i_size_read(inode);
1968 old_clusters = OCFS2_I(inode)->ip_clusters;
1970 /* communicate with ocfs2_dio_end_io */
1971 ocfs2_iocb_set_rw_locked(iocb, rw_level);
1974 ret = generic_segment_checks(iov, &nr_segs, &ocount,
1980 ret = generic_write_checks(file, ppos, &count,
1981 S_ISBLK(inode->i_mode));
1985 written = generic_file_direct_write(iocb, iov, &nr_segs, *ppos,
1986 ppos, count, ocount);
1989 * direct write may have instantiated a few
1990 * blocks outside i_size. Trim these off again.
1991 * Don't need i_size_read because we hold i_mutex.
1993 if (*ppos + count > inode->i_size)
1994 vmtruncate(inode, inode->i_size);
1999 written = generic_file_aio_write_nolock(iocb, iov, nr_segs,
2004 /* buffered aio wouldn't have proper lock coverage today */
2005 BUG_ON(ret == -EIOCBQUEUED && !(file->f_flags & O_DIRECT));
2007 if ((file->f_flags & O_SYNC && !direct_io) || IS_SYNC(inode)) {
2009 * The generic write paths have handled getting data
2010 * to disk, but since we don't make use of the dirty
2011 * inode list, a manual journal commit is necessary
2014 if (old_size != i_size_read(inode) ||
2015 old_clusters != OCFS2_I(inode)->ip_clusters) {
2016 ret = jbd2_journal_force_commit(osb->journal->j_journal);
2023 * deep in g_f_a_w_n()->ocfs2_direct_IO we pass in a ocfs2_dio_end_io
2024 * function pointer which is called when o_direct io completes so that
2025 * it can unlock our rw lock. (it's the clustered equivalent of
2026 * i_alloc_sem; protects truncate from racing with pending ios).
2027 * Unfortunately there are error cases which call end_io and others
2028 * that don't. so we don't have to unlock the rw_lock if either an
2029 * async dio is going to do it in the future or an end_io after an
2030 * error has already done it.
2032 if (ret == -EIOCBQUEUED || !ocfs2_iocb_is_rw_locked(iocb)) {
2039 ocfs2_rw_unlock(inode, rw_level);
2043 up_read(&inode->i_alloc_sem);
2045 mutex_unlock(&inode->i_mutex);
2053 static int ocfs2_splice_to_file(struct pipe_inode_info *pipe,
2055 struct splice_desc *sd)
2059 ret = ocfs2_prepare_inode_for_write(out->f_path.dentry, &sd->pos,
2060 sd->total_len, 0, NULL);
2066 return splice_from_pipe_feed(pipe, sd, pipe_to_file);
2069 static ssize_t ocfs2_file_splice_write(struct pipe_inode_info *pipe,
2076 struct address_space *mapping = out->f_mapping;
2077 struct inode *inode = mapping->host;
2078 struct splice_desc sd = {
2085 mlog_entry("(0x%p, 0x%p, %u, '%.*s')\n", out, pipe,
2087 out->f_path.dentry->d_name.len,
2088 out->f_path.dentry->d_name.name);
2091 mutex_lock_nested(&pipe->inode->i_mutex, I_MUTEX_PARENT);
2093 splice_from_pipe_begin(&sd);
2095 ret = splice_from_pipe_next(pipe, &sd);
2099 mutex_lock_nested(&inode->i_mutex, I_MUTEX_CHILD);
2100 ret = ocfs2_rw_lock(inode, 1);
2104 ret = ocfs2_splice_to_file(pipe, out, &sd);
2105 ocfs2_rw_unlock(inode, 1);
2107 mutex_unlock(&inode->i_mutex);
2109 splice_from_pipe_end(pipe, &sd);
2112 mutex_unlock(&pipe->inode->i_mutex);
2115 ret = sd.num_spliced;
2118 unsigned long nr_pages;
2121 nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
2124 * If file or inode is SYNC and we actually wrote some data,
2127 if (unlikely((out->f_flags & O_SYNC) || IS_SYNC(inode))) {
2130 mutex_lock(&inode->i_mutex);
2131 err = ocfs2_rw_lock(inode, 1);
2135 err = generic_osync_inode(inode, mapping,
2136 OSYNC_METADATA|OSYNC_DATA);
2137 ocfs2_rw_unlock(inode, 1);
2139 mutex_unlock(&inode->i_mutex);
2144 balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
2151 static ssize_t ocfs2_file_splice_read(struct file *in,
2153 struct pipe_inode_info *pipe,
2157 int ret = 0, lock_level = 0;
2158 struct inode *inode = in->f_path.dentry->d_inode;
2160 mlog_entry("(0x%p, 0x%p, %u, '%.*s')\n", in, pipe,
2162 in->f_path.dentry->d_name.len,
2163 in->f_path.dentry->d_name.name);
2166 * See the comment in ocfs2_file_aio_read()
2168 ret = ocfs2_inode_lock_atime(inode, in->f_vfsmnt, &lock_level);
2173 ocfs2_inode_unlock(inode, lock_level);
2175 ret = generic_file_splice_read(in, ppos, pipe, len, flags);
2182 static ssize_t ocfs2_file_aio_read(struct kiocb *iocb,
2183 const struct iovec *iov,
2184 unsigned long nr_segs,
2187 int ret = 0, rw_level = -1, have_alloc_sem = 0, lock_level = 0;
2188 struct file *filp = iocb->ki_filp;
2189 struct inode *inode = filp->f_path.dentry->d_inode;
2191 mlog_entry("(0x%p, %u, '%.*s')\n", filp,
2192 (unsigned int)nr_segs,
2193 filp->f_path.dentry->d_name.len,
2194 filp->f_path.dentry->d_name.name);
2203 * buffered reads protect themselves in ->readpage(). O_DIRECT reads
2204 * need locks to protect pending reads from racing with truncate.
2206 if (filp->f_flags & O_DIRECT) {
2207 down_read(&inode->i_alloc_sem);
2210 ret = ocfs2_rw_lock(inode, 0);
2216 /* communicate with ocfs2_dio_end_io */
2217 ocfs2_iocb_set_rw_locked(iocb, rw_level);
2221 * We're fine letting folks race truncates and extending
2222 * writes with read across the cluster, just like they can
2223 * locally. Hence no rw_lock during read.
2225 * Take and drop the meta data lock to update inode fields
2226 * like i_size. This allows the checks down below
2227 * generic_file_aio_read() a chance of actually working.
2229 ret = ocfs2_inode_lock_atime(inode, filp->f_vfsmnt, &lock_level);
2234 ocfs2_inode_unlock(inode, lock_level);
2236 ret = generic_file_aio_read(iocb, iov, nr_segs, iocb->ki_pos);
2238 mlog(0, "generic_file_aio_read returned -EINVAL\n");
2240 /* buffered aio wouldn't have proper lock coverage today */
2241 BUG_ON(ret == -EIOCBQUEUED && !(filp->f_flags & O_DIRECT));
2243 /* see ocfs2_file_aio_write */
2244 if (ret == -EIOCBQUEUED || !ocfs2_iocb_is_rw_locked(iocb)) {
2251 up_read(&inode->i_alloc_sem);
2253 ocfs2_rw_unlock(inode, rw_level);
2259 const struct inode_operations ocfs2_file_iops = {
2260 .setattr = ocfs2_setattr,
2261 .getattr = ocfs2_getattr,
2262 .permission = ocfs2_permission,
2263 .setxattr = generic_setxattr,
2264 .getxattr = generic_getxattr,
2265 .listxattr = ocfs2_listxattr,
2266 .removexattr = generic_removexattr,
2267 .fallocate = ocfs2_fallocate,
2268 .fiemap = ocfs2_fiemap,
2271 const struct inode_operations ocfs2_special_file_iops = {
2272 .setattr = ocfs2_setattr,
2273 .getattr = ocfs2_getattr,
2274 .permission = ocfs2_permission,
2278 * Other than ->lock, keep ocfs2_fops and ocfs2_dops in sync with
2279 * ocfs2_fops_no_plocks and ocfs2_dops_no_plocks!
2281 const struct file_operations ocfs2_fops = {
2282 .llseek = generic_file_llseek,
2283 .read = do_sync_read,
2284 .write = do_sync_write,
2286 .fsync = ocfs2_sync_file,
2287 .release = ocfs2_file_release,
2288 .open = ocfs2_file_open,
2289 .aio_read = ocfs2_file_aio_read,
2290 .aio_write = ocfs2_file_aio_write,
2291 .unlocked_ioctl = ocfs2_ioctl,
2292 #ifdef CONFIG_COMPAT
2293 .compat_ioctl = ocfs2_compat_ioctl,
2296 .flock = ocfs2_flock,
2297 .splice_read = ocfs2_file_splice_read,
2298 .splice_write = ocfs2_file_splice_write,
2301 const struct file_operations ocfs2_dops = {
2302 .llseek = generic_file_llseek,
2303 .read = generic_read_dir,
2304 .readdir = ocfs2_readdir,
2305 .fsync = ocfs2_sync_file,
2306 .release = ocfs2_dir_release,
2307 .open = ocfs2_dir_open,
2308 .unlocked_ioctl = ocfs2_ioctl,
2309 #ifdef CONFIG_COMPAT
2310 .compat_ioctl = ocfs2_compat_ioctl,
2313 .flock = ocfs2_flock,
2317 * POSIX-lockless variants of our file_operations.
2319 * These will be used if the underlying cluster stack does not support
2320 * posix file locking, if the user passes the "localflocks" mount
2321 * option, or if we have a local-only fs.
2323 * ocfs2_flock is in here because all stacks handle UNIX file locks,
2324 * so we still want it in the case of no stack support for
2325 * plocks. Internally, it will do the right thing when asked to ignore
2328 const struct file_operations ocfs2_fops_no_plocks = {
2329 .llseek = generic_file_llseek,
2330 .read = do_sync_read,
2331 .write = do_sync_write,
2333 .fsync = ocfs2_sync_file,
2334 .release = ocfs2_file_release,
2335 .open = ocfs2_file_open,
2336 .aio_read = ocfs2_file_aio_read,
2337 .aio_write = ocfs2_file_aio_write,
2338 .unlocked_ioctl = ocfs2_ioctl,
2339 #ifdef CONFIG_COMPAT
2340 .compat_ioctl = ocfs2_compat_ioctl,
2342 .flock = ocfs2_flock,
2343 .splice_read = ocfs2_file_splice_read,
2344 .splice_write = ocfs2_file_splice_write,
2347 const struct file_operations ocfs2_dops_no_plocks = {
2348 .llseek = generic_file_llseek,
2349 .read = generic_read_dir,
2350 .readdir = ocfs2_readdir,
2351 .fsync = ocfs2_sync_file,
2352 .release = ocfs2_dir_release,
2353 .open = ocfs2_dir_open,
2354 .unlocked_ioctl = ocfs2_ioctl,
2355 #ifdef CONFIG_COMPAT
2356 .compat_ioctl = ocfs2_compat_ioctl,
2358 .flock = ocfs2_flock,