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 if (file->f_mode & FMODE_WRITE)
111 dquot_initialize(inode);
113 spin_lock(&oi->ip_lock);
115 /* Check that the inode hasn't been wiped from disk by another
116 * node. If it hasn't then we're safe as long as we hold the
117 * spin lock until our increment of open count. */
118 if (OCFS2_I(inode)->ip_flags & OCFS2_INODE_DELETED) {
119 spin_unlock(&oi->ip_lock);
126 oi->ip_flags |= OCFS2_INODE_OPEN_DIRECT;
129 spin_unlock(&oi->ip_lock);
131 status = ocfs2_init_file_private(inode, file);
134 * We want to set open count back if we're failing the
137 spin_lock(&oi->ip_lock);
139 spin_unlock(&oi->ip_lock);
147 static int ocfs2_file_release(struct inode *inode, struct file *file)
149 struct ocfs2_inode_info *oi = OCFS2_I(inode);
151 mlog_entry("(0x%p, 0x%p, '%.*s')\n", inode, file,
152 file->f_path.dentry->d_name.len,
153 file->f_path.dentry->d_name.name);
155 spin_lock(&oi->ip_lock);
156 if (!--oi->ip_open_count)
157 oi->ip_flags &= ~OCFS2_INODE_OPEN_DIRECT;
158 spin_unlock(&oi->ip_lock);
160 ocfs2_free_file_private(inode, file);
167 static int ocfs2_dir_open(struct inode *inode, struct file *file)
169 return ocfs2_init_file_private(inode, file);
172 static int ocfs2_dir_release(struct inode *inode, struct file *file)
174 ocfs2_free_file_private(inode, file);
178 static int ocfs2_sync_file(struct file *file, int datasync)
182 struct dentry *dentry = file->f_path.dentry;
183 struct inode *inode = file->f_mapping->host;
184 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
186 mlog_entry("(0x%p, 0x%p, %d, '%.*s')\n", file, dentry, datasync,
187 dentry->d_name.len, dentry->d_name.name);
189 err = ocfs2_sync_inode(dentry->d_inode);
193 if (datasync && !(inode->i_state & I_DIRTY_DATASYNC))
196 journal = osb->journal->j_journal;
197 err = jbd2_journal_force_commit(journal);
202 return (err < 0) ? -EIO : 0;
205 int ocfs2_should_update_atime(struct inode *inode,
206 struct vfsmount *vfsmnt)
209 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
211 if (ocfs2_is_hard_readonly(osb) || ocfs2_is_soft_readonly(osb))
214 if ((inode->i_flags & S_NOATIME) ||
215 ((inode->i_sb->s_flags & MS_NODIRATIME) && S_ISDIR(inode->i_mode)))
219 * We can be called with no vfsmnt structure - NFSD will
222 * Note that our action here is different than touch_atime() -
223 * if we can't tell whether this is a noatime mount, then we
224 * don't know whether to trust the value of s_atime_quantum.
229 if ((vfsmnt->mnt_flags & MNT_NOATIME) ||
230 ((vfsmnt->mnt_flags & MNT_NODIRATIME) && S_ISDIR(inode->i_mode)))
233 if (vfsmnt->mnt_flags & MNT_RELATIME) {
234 if ((timespec_compare(&inode->i_atime, &inode->i_mtime) <= 0) ||
235 (timespec_compare(&inode->i_atime, &inode->i_ctime) <= 0))
242 if ((now.tv_sec - inode->i_atime.tv_sec <= osb->s_atime_quantum))
248 int ocfs2_update_inode_atime(struct inode *inode,
249 struct buffer_head *bh)
252 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
254 struct ocfs2_dinode *di = (struct ocfs2_dinode *) bh->b_data;
258 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
259 if (IS_ERR(handle)) {
260 ret = PTR_ERR(handle);
265 ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), bh,
266 OCFS2_JOURNAL_ACCESS_WRITE);
273 * Don't use ocfs2_mark_inode_dirty() here as we don't always
274 * have i_mutex to guard against concurrent changes to other
277 inode->i_atime = CURRENT_TIME;
278 di->i_atime = cpu_to_le64(inode->i_atime.tv_sec);
279 di->i_atime_nsec = cpu_to_le32(inode->i_atime.tv_nsec);
280 ocfs2_journal_dirty(handle, bh);
283 ocfs2_commit_trans(OCFS2_SB(inode->i_sb), handle);
289 static int ocfs2_set_inode_size(handle_t *handle,
291 struct buffer_head *fe_bh,
297 i_size_write(inode, new_i_size);
298 inode->i_blocks = ocfs2_inode_sector_count(inode);
299 inode->i_ctime = inode->i_mtime = CURRENT_TIME;
301 status = ocfs2_mark_inode_dirty(handle, inode, fe_bh);
312 int ocfs2_simple_size_update(struct inode *inode,
313 struct buffer_head *di_bh,
317 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
318 handle_t *handle = NULL;
320 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
321 if (IS_ERR(handle)) {
322 ret = PTR_ERR(handle);
327 ret = ocfs2_set_inode_size(handle, inode, di_bh,
332 ocfs2_commit_trans(osb, handle);
337 static int ocfs2_cow_file_pos(struct inode *inode,
338 struct buffer_head *fe_bh,
342 u32 phys, cpos = offset >> OCFS2_SB(inode->i_sb)->s_clustersize_bits;
343 unsigned int num_clusters = 0;
344 unsigned int ext_flags = 0;
347 * If the new offset is aligned to the range of the cluster, there is
348 * no space for ocfs2_zero_range_for_truncate to fill, so no need to
351 if ((offset & (OCFS2_SB(inode->i_sb)->s_clustersize - 1)) == 0)
354 status = ocfs2_get_clusters(inode, cpos, &phys,
355 &num_clusters, &ext_flags);
361 if (!(ext_flags & OCFS2_EXT_REFCOUNTED))
364 return ocfs2_refcount_cow(inode, fe_bh, cpos, 1, cpos+1);
370 static int ocfs2_orphan_for_truncate(struct ocfs2_super *osb,
372 struct buffer_head *fe_bh,
377 struct ocfs2_dinode *di;
383 * We need to CoW the cluster contains the offset if it is reflinked
384 * since we will call ocfs2_zero_range_for_truncate later which will
385 * write "0" from offset to the end of the cluster.
387 status = ocfs2_cow_file_pos(inode, fe_bh, new_i_size);
393 /* TODO: This needs to actually orphan the inode in this
396 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
397 if (IS_ERR(handle)) {
398 status = PTR_ERR(handle);
403 status = ocfs2_journal_access_di(handle, INODE_CACHE(inode), fe_bh,
404 OCFS2_JOURNAL_ACCESS_WRITE);
411 * Do this before setting i_size.
413 cluster_bytes = ocfs2_align_bytes_to_clusters(inode->i_sb, new_i_size);
414 status = ocfs2_zero_range_for_truncate(inode, handle, new_i_size,
421 i_size_write(inode, new_i_size);
422 inode->i_ctime = inode->i_mtime = CURRENT_TIME;
424 di = (struct ocfs2_dinode *) fe_bh->b_data;
425 di->i_size = cpu_to_le64(new_i_size);
426 di->i_ctime = di->i_mtime = cpu_to_le64(inode->i_ctime.tv_sec);
427 di->i_ctime_nsec = di->i_mtime_nsec = cpu_to_le32(inode->i_ctime.tv_nsec);
429 ocfs2_journal_dirty(handle, fe_bh);
432 ocfs2_commit_trans(osb, handle);
439 static int ocfs2_truncate_file(struct inode *inode,
440 struct buffer_head *di_bh,
444 struct ocfs2_dinode *fe = NULL;
445 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
447 mlog_entry("(inode = %llu, new_i_size = %llu\n",
448 (unsigned long long)OCFS2_I(inode)->ip_blkno,
449 (unsigned long long)new_i_size);
451 /* We trust di_bh because it comes from ocfs2_inode_lock(), which
452 * already validated it */
453 fe = (struct ocfs2_dinode *) di_bh->b_data;
455 mlog_bug_on_msg(le64_to_cpu(fe->i_size) != i_size_read(inode),
456 "Inode %llu, inode i_size = %lld != di "
457 "i_size = %llu, i_flags = 0x%x\n",
458 (unsigned long long)OCFS2_I(inode)->ip_blkno,
460 (unsigned long long)le64_to_cpu(fe->i_size),
461 le32_to_cpu(fe->i_flags));
463 if (new_i_size > le64_to_cpu(fe->i_size)) {
464 mlog(0, "asked to truncate file with size (%llu) to size (%llu)!\n",
465 (unsigned long long)le64_to_cpu(fe->i_size),
466 (unsigned long long)new_i_size);
472 mlog(0, "inode %llu, i_size = %llu, new_i_size = %llu\n",
473 (unsigned long long)le64_to_cpu(fe->i_blkno),
474 (unsigned long long)le64_to_cpu(fe->i_size),
475 (unsigned long long)new_i_size);
477 /* lets handle the simple truncate cases before doing any more
478 * cluster locking. */
479 if (new_i_size == le64_to_cpu(fe->i_size))
482 down_write(&OCFS2_I(inode)->ip_alloc_sem);
484 ocfs2_resv_discard(&osb->osb_la_resmap,
485 &OCFS2_I(inode)->ip_la_data_resv);
488 * The inode lock forced other nodes to sync and drop their
489 * pages, which (correctly) happens even if we have a truncate
490 * without allocation change - ocfs2 cluster sizes can be much
491 * greater than page size, so we have to truncate them
494 unmap_mapping_range(inode->i_mapping, new_i_size + PAGE_SIZE - 1, 0, 1);
495 truncate_inode_pages(inode->i_mapping, new_i_size);
497 if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
498 status = ocfs2_truncate_inline(inode, di_bh, new_i_size,
499 i_size_read(inode), 1);
503 goto bail_unlock_sem;
506 /* alright, we're going to need to do a full blown alloc size
507 * change. Orphan the inode so that recovery can complete the
508 * truncate if necessary. This does the task of marking
510 status = ocfs2_orphan_for_truncate(osb, inode, di_bh, new_i_size);
513 goto bail_unlock_sem;
516 status = ocfs2_commit_truncate(osb, inode, di_bh);
519 goto bail_unlock_sem;
522 /* TODO: orphan dir cleanup here. */
524 up_write(&OCFS2_I(inode)->ip_alloc_sem);
527 if (!status && OCFS2_I(inode)->ip_clusters == 0)
528 status = ocfs2_try_remove_refcount_tree(inode, di_bh);
535 * extend file allocation only here.
536 * we'll update all the disk stuff, and oip->alloc_size
538 * expect stuff to be locked, a transaction started and enough data /
539 * metadata reservations in the contexts.
541 * Will return -EAGAIN, and a reason if a restart is needed.
542 * If passed in, *reason will always be set, even in error.
544 int ocfs2_add_inode_data(struct ocfs2_super *osb,
549 struct buffer_head *fe_bh,
551 struct ocfs2_alloc_context *data_ac,
552 struct ocfs2_alloc_context *meta_ac,
553 enum ocfs2_alloc_restarted *reason_ret)
556 struct ocfs2_extent_tree et;
558 ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), fe_bh);
559 ret = ocfs2_add_clusters_in_btree(handle, &et, logical_offset,
560 clusters_to_add, mark_unwritten,
561 data_ac, meta_ac, reason_ret);
566 static int __ocfs2_extend_allocation(struct inode *inode, u32 logical_start,
567 u32 clusters_to_add, int mark_unwritten)
570 int restart_func = 0;
573 struct buffer_head *bh = NULL;
574 struct ocfs2_dinode *fe = NULL;
575 handle_t *handle = NULL;
576 struct ocfs2_alloc_context *data_ac = NULL;
577 struct ocfs2_alloc_context *meta_ac = NULL;
578 enum ocfs2_alloc_restarted why;
579 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
580 struct ocfs2_extent_tree et;
583 mlog_entry("(clusters_to_add = %u)\n", clusters_to_add);
586 * This function only exists for file systems which don't
589 BUG_ON(mark_unwritten && !ocfs2_sparse_alloc(osb));
591 status = ocfs2_read_inode_block(inode, &bh);
596 fe = (struct ocfs2_dinode *) bh->b_data;
599 BUG_ON(le32_to_cpu(fe->i_clusters) != OCFS2_I(inode)->ip_clusters);
601 mlog(0, "extend inode %llu, i_size = %lld, di->i_clusters = %u, "
602 "clusters_to_add = %u\n",
603 (unsigned long long)OCFS2_I(inode)->ip_blkno,
604 (long long)i_size_read(inode), le32_to_cpu(fe->i_clusters),
606 ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), bh);
607 status = ocfs2_lock_allocators(inode, &et, clusters_to_add, 0,
614 credits = ocfs2_calc_extend_credits(osb->sb, &fe->id2.i_list,
616 handle = ocfs2_start_trans(osb, credits);
617 if (IS_ERR(handle)) {
618 status = PTR_ERR(handle);
624 restarted_transaction:
625 status = dquot_alloc_space_nodirty(inode,
626 ocfs2_clusters_to_bytes(osb->sb, clusters_to_add));
631 /* reserve a write to the file entry early on - that we if we
632 * run out of credits in the allocation path, we can still
634 status = ocfs2_journal_access_di(handle, INODE_CACHE(inode), bh,
635 OCFS2_JOURNAL_ACCESS_WRITE);
641 prev_clusters = OCFS2_I(inode)->ip_clusters;
643 status = ocfs2_add_inode_data(osb,
653 if ((status < 0) && (status != -EAGAIN)) {
654 if (status != -ENOSPC)
659 ocfs2_journal_dirty(handle, bh);
661 spin_lock(&OCFS2_I(inode)->ip_lock);
662 clusters_to_add -= (OCFS2_I(inode)->ip_clusters - prev_clusters);
663 spin_unlock(&OCFS2_I(inode)->ip_lock);
664 /* Release unused quota reservation */
665 dquot_free_space(inode,
666 ocfs2_clusters_to_bytes(osb->sb, clusters_to_add));
669 if (why != RESTART_NONE && clusters_to_add) {
670 if (why == RESTART_META) {
671 mlog(0, "restarting function.\n");
675 BUG_ON(why != RESTART_TRANS);
677 mlog(0, "restarting transaction.\n");
678 /* TODO: This can be more intelligent. */
679 credits = ocfs2_calc_extend_credits(osb->sb,
682 status = ocfs2_extend_trans(handle, credits);
684 /* handle still has to be committed at
690 goto restarted_transaction;
694 mlog(0, "fe: i_clusters = %u, i_size=%llu\n",
695 le32_to_cpu(fe->i_clusters),
696 (unsigned long long)le64_to_cpu(fe->i_size));
697 mlog(0, "inode: ip_clusters=%u, i_size=%lld\n",
698 OCFS2_I(inode)->ip_clusters, (long long)i_size_read(inode));
701 if (status < 0 && did_quota)
702 dquot_free_space(inode,
703 ocfs2_clusters_to_bytes(osb->sb, clusters_to_add));
705 ocfs2_commit_trans(osb, handle);
709 ocfs2_free_alloc_context(data_ac);
713 ocfs2_free_alloc_context(meta_ac);
716 if ((!status) && restart_func) {
727 /* Some parts of this taken from generic_cont_expand, which turned out
728 * to be too fragile to do exactly what we need without us having to
729 * worry about recursive locking in ->write_begin() and ->write_end(). */
730 static int ocfs2_write_zero_page(struct inode *inode,
733 struct address_space *mapping = inode->i_mapping;
737 handle_t *handle = NULL;
740 offset = (size & (PAGE_CACHE_SIZE-1)); /* Within page */
741 /* ugh. in prepare/commit_write, if from==to==start of block, we
742 ** skip the prepare. make sure we never send an offset for the start
745 if ((offset & (inode->i_sb->s_blocksize - 1)) == 0) {
748 index = size >> PAGE_CACHE_SHIFT;
750 page = grab_cache_page(mapping, index);
757 ret = ocfs2_prepare_write_nolock(inode, page, offset, offset);
763 if (ocfs2_should_order_data(inode)) {
764 handle = ocfs2_start_walk_page_trans(inode, page, offset,
766 if (IS_ERR(handle)) {
767 ret = PTR_ERR(handle);
773 /* must not update i_size! */
774 ret = block_commit_write(page, offset, offset);
781 ocfs2_commit_trans(OCFS2_SB(inode->i_sb), handle);
784 page_cache_release(page);
789 static int ocfs2_zero_extend(struct inode *inode,
794 struct super_block *sb = inode->i_sb;
796 start_off = ocfs2_align_bytes_to_blocks(sb, i_size_read(inode));
797 while (start_off < zero_to_size) {
798 ret = ocfs2_write_zero_page(inode, start_off);
804 start_off += sb->s_blocksize;
807 * Very large extends have the potential to lock up
808 * the cpu for extended periods of time.
817 int ocfs2_extend_no_holes(struct inode *inode, u64 new_i_size, u64 zero_to)
821 struct ocfs2_inode_info *oi = OCFS2_I(inode);
823 clusters_to_add = ocfs2_clusters_for_bytes(inode->i_sb, new_i_size);
824 if (clusters_to_add < oi->ip_clusters)
827 clusters_to_add -= oi->ip_clusters;
829 if (clusters_to_add) {
830 ret = __ocfs2_extend_allocation(inode, oi->ip_clusters,
839 * Call this even if we don't add any clusters to the tree. We
840 * still need to zero the area between the old i_size and the
843 ret = ocfs2_zero_extend(inode, zero_to);
851 static int ocfs2_extend_file(struct inode *inode,
852 struct buffer_head *di_bh,
856 struct ocfs2_inode_info *oi = OCFS2_I(inode);
860 /* setattr sometimes calls us like this. */
864 if (i_size_read(inode) == new_i_size)
866 BUG_ON(new_i_size < i_size_read(inode));
869 * Fall through for converting inline data, even if the fs
870 * supports sparse files.
872 * The check for inline data here is legal - nobody can add
873 * the feature since we have i_mutex. We must check it again
874 * after acquiring ip_alloc_sem though, as paths like mmap
875 * might have raced us to converting the inode to extents.
877 if (!(oi->ip_dyn_features & OCFS2_INLINE_DATA_FL)
878 && ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)))
879 goto out_update_size;
882 * The alloc sem blocks people in read/write from reading our
883 * allocation until we're done changing it. We depend on
884 * i_mutex to block other extend/truncate calls while we're
887 down_write(&oi->ip_alloc_sem);
889 if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
891 * We can optimize small extends by keeping the inodes
894 if (ocfs2_size_fits_inline_data(di_bh, new_i_size)) {
895 up_write(&oi->ip_alloc_sem);
896 goto out_update_size;
899 ret = ocfs2_convert_inline_data_to_extents(inode, di_bh);
901 up_write(&oi->ip_alloc_sem);
908 if (!ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)))
909 ret = ocfs2_extend_no_holes(inode, new_i_size, new_i_size);
911 up_write(&oi->ip_alloc_sem);
919 ret = ocfs2_simple_size_update(inode, di_bh, new_i_size);
927 int ocfs2_setattr(struct dentry *dentry, struct iattr *attr)
929 int status = 0, size_change;
930 struct inode *inode = dentry->d_inode;
931 struct super_block *sb = inode->i_sb;
932 struct ocfs2_super *osb = OCFS2_SB(sb);
933 struct buffer_head *bh = NULL;
934 handle_t *handle = NULL;
935 struct dquot *transfer_to[MAXQUOTAS] = { };
938 mlog_entry("(0x%p, '%.*s')\n", dentry,
939 dentry->d_name.len, dentry->d_name.name);
941 /* ensuring we don't even attempt to truncate a symlink */
942 if (S_ISLNK(inode->i_mode))
943 attr->ia_valid &= ~ATTR_SIZE;
945 if (attr->ia_valid & ATTR_MODE)
946 mlog(0, "mode change: %d\n", attr->ia_mode);
947 if (attr->ia_valid & ATTR_UID)
948 mlog(0, "uid change: %d\n", attr->ia_uid);
949 if (attr->ia_valid & ATTR_GID)
950 mlog(0, "gid change: %d\n", attr->ia_gid);
951 if (attr->ia_valid & ATTR_SIZE)
952 mlog(0, "size change...\n");
953 if (attr->ia_valid & (ATTR_ATIME | ATTR_MTIME | ATTR_CTIME))
954 mlog(0, "time change...\n");
956 #define OCFS2_VALID_ATTRS (ATTR_ATIME | ATTR_MTIME | ATTR_CTIME | ATTR_SIZE \
957 | ATTR_GID | ATTR_UID | ATTR_MODE)
958 if (!(attr->ia_valid & OCFS2_VALID_ATTRS)) {
959 mlog(0, "can't handle attrs: 0x%x\n", attr->ia_valid);
963 status = inode_change_ok(inode, attr);
967 if (is_quota_modification(inode, attr))
968 dquot_initialize(inode);
969 size_change = S_ISREG(inode->i_mode) && attr->ia_valid & ATTR_SIZE;
971 status = ocfs2_rw_lock(inode, 1);
978 status = ocfs2_inode_lock(inode, &bh, 1);
980 if (status != -ENOENT)
985 if (size_change && attr->ia_size != i_size_read(inode)) {
986 status = inode_newsize_ok(inode, attr->ia_size);
990 if (i_size_read(inode) > attr->ia_size) {
991 if (ocfs2_should_order_data(inode)) {
992 status = ocfs2_begin_ordered_truncate(inode,
997 status = ocfs2_truncate_file(inode, bh, attr->ia_size);
999 status = ocfs2_extend_file(inode, bh, attr->ia_size);
1001 if (status != -ENOSPC)
1008 if ((attr->ia_valid & ATTR_UID && attr->ia_uid != inode->i_uid) ||
1009 (attr->ia_valid & ATTR_GID && attr->ia_gid != inode->i_gid)) {
1011 * Gather pointers to quota structures so that allocation /
1012 * freeing of quota structures happens here and not inside
1013 * dquot_transfer() where we have problems with lock ordering
1015 if (attr->ia_valid & ATTR_UID && attr->ia_uid != inode->i_uid
1016 && OCFS2_HAS_RO_COMPAT_FEATURE(sb,
1017 OCFS2_FEATURE_RO_COMPAT_USRQUOTA)) {
1018 transfer_to[USRQUOTA] = dqget(sb, attr->ia_uid,
1020 if (!transfer_to[USRQUOTA]) {
1025 if (attr->ia_valid & ATTR_GID && attr->ia_gid != inode->i_gid
1026 && OCFS2_HAS_RO_COMPAT_FEATURE(sb,
1027 OCFS2_FEATURE_RO_COMPAT_GRPQUOTA)) {
1028 transfer_to[GRPQUOTA] = dqget(sb, attr->ia_gid,
1030 if (!transfer_to[GRPQUOTA]) {
1035 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS +
1036 2 * ocfs2_quota_trans_credits(sb));
1037 if (IS_ERR(handle)) {
1038 status = PTR_ERR(handle);
1042 status = __dquot_transfer(inode, transfer_to);
1046 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1047 if (IS_ERR(handle)) {
1048 status = PTR_ERR(handle);
1055 * This will intentionally not wind up calling simple_setsize(),
1056 * since all the work for a size change has been done above.
1057 * Otherwise, we could get into problems with truncate as
1058 * ip_alloc_sem is used there to protect against i_size
1061 status = inode_setattr(inode, attr);
1067 status = ocfs2_mark_inode_dirty(handle, inode, bh);
1072 ocfs2_commit_trans(osb, handle);
1074 ocfs2_inode_unlock(inode, 1);
1077 ocfs2_rw_unlock(inode, 1);
1081 /* Release quota pointers in case we acquired them */
1082 for (qtype = 0; qtype < MAXQUOTAS; qtype++)
1083 dqput(transfer_to[qtype]);
1085 if (!status && attr->ia_valid & ATTR_MODE) {
1086 status = ocfs2_acl_chmod(inode);
1095 int ocfs2_getattr(struct vfsmount *mnt,
1096 struct dentry *dentry,
1099 struct inode *inode = dentry->d_inode;
1100 struct super_block *sb = dentry->d_inode->i_sb;
1101 struct ocfs2_super *osb = sb->s_fs_info;
1106 err = ocfs2_inode_revalidate(dentry);
1113 generic_fillattr(inode, stat);
1115 /* We set the blksize from the cluster size for performance */
1116 stat->blksize = osb->s_clustersize;
1124 int ocfs2_permission(struct inode *inode, int mask)
1130 ret = ocfs2_inode_lock(inode, NULL, 0);
1137 ret = generic_permission(inode, mask, ocfs2_check_acl);
1139 ocfs2_inode_unlock(inode, 0);
1145 static int __ocfs2_write_remove_suid(struct inode *inode,
1146 struct buffer_head *bh)
1150 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1151 struct ocfs2_dinode *di;
1153 mlog_entry("(Inode %llu, mode 0%o)\n",
1154 (unsigned long long)OCFS2_I(inode)->ip_blkno, inode->i_mode);
1156 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1157 if (IS_ERR(handle)) {
1158 ret = PTR_ERR(handle);
1163 ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), bh,
1164 OCFS2_JOURNAL_ACCESS_WRITE);
1170 inode->i_mode &= ~S_ISUID;
1171 if ((inode->i_mode & S_ISGID) && (inode->i_mode & S_IXGRP))
1172 inode->i_mode &= ~S_ISGID;
1174 di = (struct ocfs2_dinode *) bh->b_data;
1175 di->i_mode = cpu_to_le16(inode->i_mode);
1177 ocfs2_journal_dirty(handle, bh);
1180 ocfs2_commit_trans(osb, handle);
1187 * Will look for holes and unwritten extents in the range starting at
1188 * pos for count bytes (inclusive).
1190 static int ocfs2_check_range_for_holes(struct inode *inode, loff_t pos,
1194 unsigned int extent_flags;
1195 u32 cpos, clusters, extent_len, phys_cpos;
1196 struct super_block *sb = inode->i_sb;
1198 cpos = pos >> OCFS2_SB(sb)->s_clustersize_bits;
1199 clusters = ocfs2_clusters_for_bytes(sb, pos + count) - cpos;
1202 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos, &extent_len,
1209 if (phys_cpos == 0 || (extent_flags & OCFS2_EXT_UNWRITTEN)) {
1214 if (extent_len > clusters)
1215 extent_len = clusters;
1217 clusters -= extent_len;
1224 static int ocfs2_write_remove_suid(struct inode *inode)
1227 struct buffer_head *bh = NULL;
1229 ret = ocfs2_read_inode_block(inode, &bh);
1235 ret = __ocfs2_write_remove_suid(inode, bh);
1242 * Allocate enough extents to cover the region starting at byte offset
1243 * start for len bytes. Existing extents are skipped, any extents
1244 * added are marked as "unwritten".
1246 static int ocfs2_allocate_unwritten_extents(struct inode *inode,
1250 u32 cpos, phys_cpos, clusters, alloc_size;
1251 u64 end = start + len;
1252 struct buffer_head *di_bh = NULL;
1254 if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1255 ret = ocfs2_read_inode_block(inode, &di_bh);
1262 * Nothing to do if the requested reservation range
1263 * fits within the inode.
1265 if (ocfs2_size_fits_inline_data(di_bh, end))
1268 ret = ocfs2_convert_inline_data_to_extents(inode, di_bh);
1276 * We consider both start and len to be inclusive.
1278 cpos = start >> OCFS2_SB(inode->i_sb)->s_clustersize_bits;
1279 clusters = ocfs2_clusters_for_bytes(inode->i_sb, start + len);
1283 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos,
1291 * Hole or existing extent len can be arbitrary, so
1292 * cap it to our own allocation request.
1294 if (alloc_size > clusters)
1295 alloc_size = clusters;
1299 * We already have an allocation at this
1300 * region so we can safely skip it.
1305 ret = __ocfs2_extend_allocation(inode, cpos, alloc_size, 1);
1314 clusters -= alloc_size;
1325 * Truncate a byte range, avoiding pages within partial clusters. This
1326 * preserves those pages for the zeroing code to write to.
1328 static void ocfs2_truncate_cluster_pages(struct inode *inode, u64 byte_start,
1331 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1333 struct address_space *mapping = inode->i_mapping;
1335 start = (loff_t)ocfs2_align_bytes_to_clusters(inode->i_sb, byte_start);
1336 end = byte_start + byte_len;
1337 end = end & ~(osb->s_clustersize - 1);
1340 unmap_mapping_range(mapping, start, end - start, 0);
1341 truncate_inode_pages_range(mapping, start, end - 1);
1345 static int ocfs2_zero_partial_clusters(struct inode *inode,
1349 u64 tmpend, end = start + len;
1350 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1351 unsigned int csize = osb->s_clustersize;
1355 * The "start" and "end" values are NOT necessarily part of
1356 * the range whose allocation is being deleted. Rather, this
1357 * is what the user passed in with the request. We must zero
1358 * partial clusters here. There's no need to worry about
1359 * physical allocation - the zeroing code knows to skip holes.
1361 mlog(0, "byte start: %llu, end: %llu\n",
1362 (unsigned long long)start, (unsigned long long)end);
1365 * If both edges are on a cluster boundary then there's no
1366 * zeroing required as the region is part of the allocation to
1369 if ((start & (csize - 1)) == 0 && (end & (csize - 1)) == 0)
1372 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1373 if (IS_ERR(handle)) {
1374 ret = PTR_ERR(handle);
1380 * We want to get the byte offset of the end of the 1st cluster.
1382 tmpend = (u64)osb->s_clustersize + (start & ~(osb->s_clustersize - 1));
1386 mlog(0, "1st range: start: %llu, tmpend: %llu\n",
1387 (unsigned long long)start, (unsigned long long)tmpend);
1389 ret = ocfs2_zero_range_for_truncate(inode, handle, start, tmpend);
1395 * This may make start and end equal, but the zeroing
1396 * code will skip any work in that case so there's no
1397 * need to catch it up here.
1399 start = end & ~(osb->s_clustersize - 1);
1401 mlog(0, "2nd range: start: %llu, end: %llu\n",
1402 (unsigned long long)start, (unsigned long long)end);
1404 ret = ocfs2_zero_range_for_truncate(inode, handle, start, end);
1409 ocfs2_commit_trans(osb, handle);
1414 static int ocfs2_find_rec(struct ocfs2_extent_list *el, u32 pos)
1417 struct ocfs2_extent_rec *rec = NULL;
1419 for (i = le16_to_cpu(el->l_next_free_rec) - 1; i >= 0; i--) {
1421 rec = &el->l_recs[i];
1423 if (le32_to_cpu(rec->e_cpos) < pos)
1431 * Helper to calculate the punching pos and length in one run, we handle the
1432 * following three cases in order:
1434 * - remove the entire record
1435 * - remove a partial record
1436 * - no record needs to be removed (hole-punching completed)
1438 static void ocfs2_calc_trunc_pos(struct inode *inode,
1439 struct ocfs2_extent_list *el,
1440 struct ocfs2_extent_rec *rec,
1441 u32 trunc_start, u32 *trunc_cpos,
1442 u32 *trunc_len, u32 *trunc_end,
1443 u64 *blkno, int *done)
1448 range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
1450 if (le32_to_cpu(rec->e_cpos) >= trunc_start) {
1451 *trunc_cpos = le32_to_cpu(rec->e_cpos);
1453 * Skip holes if any.
1455 if (range < *trunc_end)
1457 *trunc_len = *trunc_end - le32_to_cpu(rec->e_cpos);
1458 *blkno = le64_to_cpu(rec->e_blkno);
1459 *trunc_end = le32_to_cpu(rec->e_cpos);
1460 } else if (range > trunc_start) {
1461 *trunc_cpos = trunc_start;
1462 *trunc_len = *trunc_end - trunc_start;
1463 coff = trunc_start - le32_to_cpu(rec->e_cpos);
1464 *blkno = le64_to_cpu(rec->e_blkno) +
1465 ocfs2_clusters_to_blocks(inode->i_sb, coff);
1466 *trunc_end = trunc_start;
1469 * It may have two following possibilities:
1471 * - last record has been removed
1472 * - trunc_start was within a hole
1474 * both two cases mean the completion of hole punching.
1482 static int ocfs2_remove_inode_range(struct inode *inode,
1483 struct buffer_head *di_bh, u64 byte_start,
1486 int ret = 0, flags = 0, done = 0, i;
1487 u32 trunc_start, trunc_len, trunc_end, trunc_cpos, phys_cpos;
1489 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1490 struct ocfs2_cached_dealloc_ctxt dealloc;
1491 struct address_space *mapping = inode->i_mapping;
1492 struct ocfs2_extent_tree et;
1493 struct ocfs2_path *path = NULL;
1494 struct ocfs2_extent_list *el = NULL;
1495 struct ocfs2_extent_rec *rec = NULL;
1496 struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
1497 u64 blkno, refcount_loc = le64_to_cpu(di->i_refcount_loc);
1499 ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), di_bh);
1500 ocfs2_init_dealloc_ctxt(&dealloc);
1505 if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1506 ret = ocfs2_truncate_inline(inode, di_bh, byte_start,
1507 byte_start + byte_len, 0);
1513 * There's no need to get fancy with the page cache
1514 * truncate of an inline-data inode. We're talking
1515 * about less than a page here, which will be cached
1516 * in the dinode buffer anyway.
1518 unmap_mapping_range(mapping, 0, 0, 0);
1519 truncate_inode_pages(mapping, 0);
1524 * For reflinks, we may need to CoW 2 clusters which might be
1525 * partially zero'd later, if hole's start and end offset were
1526 * within one cluster(means is not exactly aligned to clustersize).
1529 if (OCFS2_I(inode)->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL) {
1531 ret = ocfs2_cow_file_pos(inode, di_bh, byte_start);
1537 ret = ocfs2_cow_file_pos(inode, di_bh, byte_start + byte_len);
1544 trunc_start = ocfs2_clusters_for_bytes(osb->sb, byte_start);
1545 trunc_end = (byte_start + byte_len) >> osb->s_clustersize_bits;
1546 cluster_in_el = trunc_end;
1548 mlog(0, "Inode: %llu, start: %llu, len: %llu, cstart: %u, cend: %u\n",
1549 (unsigned long long)OCFS2_I(inode)->ip_blkno,
1550 (unsigned long long)byte_start,
1551 (unsigned long long)byte_len, trunc_start, trunc_end);
1553 ret = ocfs2_zero_partial_clusters(inode, byte_start, byte_len);
1559 path = ocfs2_new_path_from_et(&et);
1566 while (trunc_end > trunc_start) {
1568 ret = ocfs2_find_path(INODE_CACHE(inode), path,
1575 el = path_leaf_el(path);
1577 i = ocfs2_find_rec(el, trunc_end);
1579 * Need to go to previous extent block.
1582 if (path->p_tree_depth == 0)
1585 ret = ocfs2_find_cpos_for_left_leaf(inode->i_sb,
1594 * We've reached the leftmost extent block,
1595 * it's safe to leave.
1597 if (cluster_in_el == 0)
1601 * The 'pos' searched for previous extent block is
1602 * always one cluster less than actual trunc_end.
1604 trunc_end = cluster_in_el + 1;
1606 ocfs2_reinit_path(path, 1);
1611 rec = &el->l_recs[i];
1613 ocfs2_calc_trunc_pos(inode, el, rec, trunc_start, &trunc_cpos,
1614 &trunc_len, &trunc_end, &blkno, &done);
1618 flags = rec->e_flags;
1619 phys_cpos = ocfs2_blocks_to_clusters(inode->i_sb, blkno);
1621 ret = ocfs2_remove_btree_range(inode, &et, trunc_cpos,
1622 phys_cpos, trunc_len, flags,
1623 &dealloc, refcount_loc);
1629 cluster_in_el = trunc_end;
1631 ocfs2_reinit_path(path, 1);
1634 ocfs2_truncate_cluster_pages(inode, byte_start, byte_len);
1637 ocfs2_schedule_truncate_log_flush(osb, 1);
1638 ocfs2_run_deallocs(osb, &dealloc);
1644 * Parts of this function taken from xfs_change_file_space()
1646 static int __ocfs2_change_file_space(struct file *file, struct inode *inode,
1647 loff_t f_pos, unsigned int cmd,
1648 struct ocfs2_space_resv *sr,
1654 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1655 struct buffer_head *di_bh = NULL;
1657 unsigned long long max_off = inode->i_sb->s_maxbytes;
1659 if (ocfs2_is_hard_readonly(osb) || ocfs2_is_soft_readonly(osb))
1662 mutex_lock(&inode->i_mutex);
1665 * This prevents concurrent writes on other nodes
1667 ret = ocfs2_rw_lock(inode, 1);
1673 ret = ocfs2_inode_lock(inode, &di_bh, 1);
1679 if (inode->i_flags & (S_IMMUTABLE|S_APPEND)) {
1681 goto out_inode_unlock;
1684 switch (sr->l_whence) {
1685 case 0: /*SEEK_SET*/
1687 case 1: /*SEEK_CUR*/
1688 sr->l_start += f_pos;
1690 case 2: /*SEEK_END*/
1691 sr->l_start += i_size_read(inode);
1695 goto out_inode_unlock;
1699 llen = sr->l_len > 0 ? sr->l_len - 1 : sr->l_len;
1702 || sr->l_start > max_off
1703 || (sr->l_start + llen) < 0
1704 || (sr->l_start + llen) > max_off) {
1706 goto out_inode_unlock;
1708 size = sr->l_start + sr->l_len;
1710 if (cmd == OCFS2_IOC_RESVSP || cmd == OCFS2_IOC_RESVSP64) {
1711 if (sr->l_len <= 0) {
1713 goto out_inode_unlock;
1717 if (file && should_remove_suid(file->f_path.dentry)) {
1718 ret = __ocfs2_write_remove_suid(inode, di_bh);
1721 goto out_inode_unlock;
1725 down_write(&OCFS2_I(inode)->ip_alloc_sem);
1727 case OCFS2_IOC_RESVSP:
1728 case OCFS2_IOC_RESVSP64:
1730 * This takes unsigned offsets, but the signed ones we
1731 * pass have been checked against overflow above.
1733 ret = ocfs2_allocate_unwritten_extents(inode, sr->l_start,
1736 case OCFS2_IOC_UNRESVSP:
1737 case OCFS2_IOC_UNRESVSP64:
1738 ret = ocfs2_remove_inode_range(inode, di_bh, sr->l_start,
1744 up_write(&OCFS2_I(inode)->ip_alloc_sem);
1747 goto out_inode_unlock;
1751 * We update c/mtime for these changes
1753 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1754 if (IS_ERR(handle)) {
1755 ret = PTR_ERR(handle);
1757 goto out_inode_unlock;
1760 if (change_size && i_size_read(inode) < size)
1761 i_size_write(inode, size);
1763 inode->i_ctime = inode->i_mtime = CURRENT_TIME;
1764 ret = ocfs2_mark_inode_dirty(handle, inode, di_bh);
1768 ocfs2_commit_trans(osb, handle);
1772 ocfs2_inode_unlock(inode, 1);
1774 ocfs2_rw_unlock(inode, 1);
1777 mutex_unlock(&inode->i_mutex);
1781 int ocfs2_change_file_space(struct file *file, unsigned int cmd,
1782 struct ocfs2_space_resv *sr)
1784 struct inode *inode = file->f_path.dentry->d_inode;
1785 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1787 if ((cmd == OCFS2_IOC_RESVSP || cmd == OCFS2_IOC_RESVSP64) &&
1788 !ocfs2_writes_unwritten_extents(osb))
1790 else if ((cmd == OCFS2_IOC_UNRESVSP || cmd == OCFS2_IOC_UNRESVSP64) &&
1791 !ocfs2_sparse_alloc(osb))
1794 if (!S_ISREG(inode->i_mode))
1797 if (!(file->f_mode & FMODE_WRITE))
1800 return __ocfs2_change_file_space(file, inode, file->f_pos, cmd, sr, 0);
1803 static long ocfs2_fallocate(struct inode *inode, int mode, loff_t offset,
1806 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1807 struct ocfs2_space_resv sr;
1808 int change_size = 1;
1810 if (!ocfs2_writes_unwritten_extents(osb))
1813 if (S_ISDIR(inode->i_mode))
1816 if (mode & FALLOC_FL_KEEP_SIZE)
1820 sr.l_start = (s64)offset;
1821 sr.l_len = (s64)len;
1823 return __ocfs2_change_file_space(NULL, inode, offset,
1824 OCFS2_IOC_RESVSP64, &sr, change_size);
1827 int ocfs2_check_range_for_refcount(struct inode *inode, loff_t pos,
1831 unsigned int extent_flags;
1832 u32 cpos, clusters, extent_len, phys_cpos;
1833 struct super_block *sb = inode->i_sb;
1835 if (!ocfs2_refcount_tree(OCFS2_SB(inode->i_sb)) ||
1836 !(OCFS2_I(inode)->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL) ||
1837 OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
1840 cpos = pos >> OCFS2_SB(sb)->s_clustersize_bits;
1841 clusters = ocfs2_clusters_for_bytes(sb, pos + count) - cpos;
1844 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos, &extent_len,
1851 if (phys_cpos && (extent_flags & OCFS2_EXT_REFCOUNTED)) {
1856 if (extent_len > clusters)
1857 extent_len = clusters;
1859 clusters -= extent_len;
1866 static int ocfs2_prepare_inode_for_refcount(struct inode *inode,
1867 loff_t pos, size_t count,
1871 struct buffer_head *di_bh = NULL;
1872 u32 cpos = pos >> OCFS2_SB(inode->i_sb)->s_clustersize_bits;
1874 ocfs2_clusters_for_bytes(inode->i_sb, pos + count) - cpos;
1876 ret = ocfs2_inode_lock(inode, &di_bh, 1);
1884 ret = ocfs2_refcount_cow(inode, di_bh, cpos, clusters, UINT_MAX);
1892 static int ocfs2_prepare_inode_for_write(struct dentry *dentry,
1899 int ret = 0, meta_level = 0;
1900 struct inode *inode = dentry->d_inode;
1901 loff_t saved_pos, end;
1904 * We start with a read level meta lock and only jump to an ex
1905 * if we need to make modifications here.
1908 ret = ocfs2_inode_lock(inode, NULL, meta_level);
1915 /* Clear suid / sgid if necessary. We do this here
1916 * instead of later in the write path because
1917 * remove_suid() calls ->setattr without any hint that
1918 * we may have already done our cluster locking. Since
1919 * ocfs2_setattr() *must* take cluster locks to
1920 * proceeed, this will lead us to recursively lock the
1921 * inode. There's also the dinode i_size state which
1922 * can be lost via setattr during extending writes (we
1923 * set inode->i_size at the end of a write. */
1924 if (should_remove_suid(dentry)) {
1925 if (meta_level == 0) {
1926 ocfs2_inode_unlock(inode, meta_level);
1931 ret = ocfs2_write_remove_suid(inode);
1938 /* work on a copy of ppos until we're sure that we won't have
1939 * to recalculate it due to relocking. */
1941 saved_pos = i_size_read(inode);
1942 mlog(0, "O_APPEND: inode->i_size=%llu\n", saved_pos);
1947 end = saved_pos + count;
1949 ret = ocfs2_check_range_for_refcount(inode, saved_pos, count);
1951 ocfs2_inode_unlock(inode, meta_level);
1954 ret = ocfs2_prepare_inode_for_refcount(inode,
1970 * Skip the O_DIRECT checks if we don't need
1973 if (!direct_io || !(*direct_io))
1977 * There's no sane way to do direct writes to an inode
1980 if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1986 * Allowing concurrent direct writes means
1987 * i_size changes wouldn't be synchronized, so
1988 * one node could wind up truncating another
1991 if (end > i_size_read(inode)) {
1997 * We don't fill holes during direct io, so
1998 * check for them here. If any are found, the
1999 * caller will have to retake some cluster
2000 * locks and initiate the io as buffered.
2002 ret = ocfs2_check_range_for_holes(inode, saved_pos, count);
2015 if (meta_level >= 0)
2016 ocfs2_inode_unlock(inode, meta_level);
2022 static ssize_t ocfs2_file_aio_write(struct kiocb *iocb,
2023 const struct iovec *iov,
2024 unsigned long nr_segs,
2027 int ret, direct_io, appending, rw_level, have_alloc_sem = 0;
2028 int can_do_direct, has_refcount = 0;
2029 ssize_t written = 0;
2030 size_t ocount; /* original count */
2031 size_t count; /* after file limit checks */
2032 loff_t old_size, *ppos = &iocb->ki_pos;
2034 struct file *file = iocb->ki_filp;
2035 struct inode *inode = file->f_path.dentry->d_inode;
2036 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
2038 mlog_entry("(0x%p, %u, '%.*s')\n", file,
2039 (unsigned int)nr_segs,
2040 file->f_path.dentry->d_name.len,
2041 file->f_path.dentry->d_name.name);
2043 if (iocb->ki_left == 0)
2046 vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
2048 appending = file->f_flags & O_APPEND ? 1 : 0;
2049 direct_io = file->f_flags & O_DIRECT ? 1 : 0;
2051 mutex_lock(&inode->i_mutex);
2054 /* to match setattr's i_mutex -> i_alloc_sem -> rw_lock ordering */
2056 down_read(&inode->i_alloc_sem);
2060 /* concurrent O_DIRECT writes are allowed */
2061 rw_level = !direct_io;
2062 ret = ocfs2_rw_lock(inode, rw_level);
2068 can_do_direct = direct_io;
2069 ret = ocfs2_prepare_inode_for_write(file->f_path.dentry, ppos,
2070 iocb->ki_left, appending,
2071 &can_do_direct, &has_refcount);
2078 * We can't complete the direct I/O as requested, fall back to
2081 if (direct_io && !can_do_direct) {
2082 ocfs2_rw_unlock(inode, rw_level);
2083 up_read(&inode->i_alloc_sem);
2093 * To later detect whether a journal commit for sync writes is
2094 * necessary, we sample i_size, and cluster count here.
2096 old_size = i_size_read(inode);
2097 old_clusters = OCFS2_I(inode)->ip_clusters;
2099 /* communicate with ocfs2_dio_end_io */
2100 ocfs2_iocb_set_rw_locked(iocb, rw_level);
2102 ret = generic_segment_checks(iov, &nr_segs, &ocount,
2108 ret = generic_write_checks(file, ppos, &count,
2109 S_ISBLK(inode->i_mode));
2114 written = generic_file_direct_write(iocb, iov, &nr_segs, *ppos,
2115 ppos, count, ocount);
2118 * direct write may have instantiated a few
2119 * blocks outside i_size. Trim these off again.
2120 * Don't need i_size_read because we hold i_mutex.
2122 * XXX(hch): this looks buggy because ocfs2 did not
2123 * actually implement ->truncate. Take a look at
2124 * the new truncate sequence and update this accordingly
2126 if (*ppos + count > inode->i_size)
2127 simple_setsize(inode, inode->i_size);
2132 current->backing_dev_info = file->f_mapping->backing_dev_info;
2133 written = generic_file_buffered_write(iocb, iov, nr_segs, *ppos,
2135 current->backing_dev_info = NULL;
2139 /* buffered aio wouldn't have proper lock coverage today */
2140 BUG_ON(ret == -EIOCBQUEUED && !(file->f_flags & O_DIRECT));
2142 if (((file->f_flags & O_DSYNC) && !direct_io) || IS_SYNC(inode) ||
2143 ((file->f_flags & O_DIRECT) && has_refcount)) {
2144 ret = filemap_fdatawrite_range(file->f_mapping, pos,
2149 if (!ret && ((old_size != i_size_read(inode)) ||
2150 (old_clusters != OCFS2_I(inode)->ip_clusters) ||
2152 ret = jbd2_journal_force_commit(osb->journal->j_journal);
2158 ret = filemap_fdatawait_range(file->f_mapping, pos,
2163 * deep in g_f_a_w_n()->ocfs2_direct_IO we pass in a ocfs2_dio_end_io
2164 * function pointer which is called when o_direct io completes so that
2165 * it can unlock our rw lock. (it's the clustered equivalent of
2166 * i_alloc_sem; protects truncate from racing with pending ios).
2167 * Unfortunately there are error cases which call end_io and others
2168 * that don't. so we don't have to unlock the rw_lock if either an
2169 * async dio is going to do it in the future or an end_io after an
2170 * error has already done it.
2172 if ((ret == -EIOCBQUEUED) || (!ocfs2_iocb_is_rw_locked(iocb))) {
2179 ocfs2_rw_unlock(inode, rw_level);
2183 up_read(&inode->i_alloc_sem);
2185 mutex_unlock(&inode->i_mutex);
2193 static int ocfs2_splice_to_file(struct pipe_inode_info *pipe,
2195 struct splice_desc *sd)
2199 ret = ocfs2_prepare_inode_for_write(out->f_path.dentry, &sd->pos,
2200 sd->total_len, 0, NULL, NULL);
2206 return splice_from_pipe_feed(pipe, sd, pipe_to_file);
2209 static ssize_t ocfs2_file_splice_write(struct pipe_inode_info *pipe,
2216 struct address_space *mapping = out->f_mapping;
2217 struct inode *inode = mapping->host;
2218 struct splice_desc sd = {
2225 mlog_entry("(0x%p, 0x%p, %u, '%.*s')\n", out, pipe,
2227 out->f_path.dentry->d_name.len,
2228 out->f_path.dentry->d_name.name);
2231 mutex_lock_nested(&pipe->inode->i_mutex, I_MUTEX_PARENT);
2233 splice_from_pipe_begin(&sd);
2235 ret = splice_from_pipe_next(pipe, &sd);
2239 mutex_lock_nested(&inode->i_mutex, I_MUTEX_CHILD);
2240 ret = ocfs2_rw_lock(inode, 1);
2244 ret = ocfs2_splice_to_file(pipe, out, &sd);
2245 ocfs2_rw_unlock(inode, 1);
2247 mutex_unlock(&inode->i_mutex);
2249 splice_from_pipe_end(pipe, &sd);
2252 mutex_unlock(&pipe->inode->i_mutex);
2255 ret = sd.num_spliced;
2258 unsigned long nr_pages;
2261 nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
2263 err = generic_write_sync(out, *ppos, ret);
2269 balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
2276 static ssize_t ocfs2_file_splice_read(struct file *in,
2278 struct pipe_inode_info *pipe,
2282 int ret = 0, lock_level = 0;
2283 struct inode *inode = in->f_path.dentry->d_inode;
2285 mlog_entry("(0x%p, 0x%p, %u, '%.*s')\n", in, pipe,
2287 in->f_path.dentry->d_name.len,
2288 in->f_path.dentry->d_name.name);
2291 * See the comment in ocfs2_file_aio_read()
2293 ret = ocfs2_inode_lock_atime(inode, in->f_vfsmnt, &lock_level);
2298 ocfs2_inode_unlock(inode, lock_level);
2300 ret = generic_file_splice_read(in, ppos, pipe, len, flags);
2307 static ssize_t ocfs2_file_aio_read(struct kiocb *iocb,
2308 const struct iovec *iov,
2309 unsigned long nr_segs,
2312 int ret = 0, rw_level = -1, have_alloc_sem = 0, lock_level = 0;
2313 struct file *filp = iocb->ki_filp;
2314 struct inode *inode = filp->f_path.dentry->d_inode;
2316 mlog_entry("(0x%p, %u, '%.*s')\n", filp,
2317 (unsigned int)nr_segs,
2318 filp->f_path.dentry->d_name.len,
2319 filp->f_path.dentry->d_name.name);
2328 * buffered reads protect themselves in ->readpage(). O_DIRECT reads
2329 * need locks to protect pending reads from racing with truncate.
2331 if (filp->f_flags & O_DIRECT) {
2332 down_read(&inode->i_alloc_sem);
2335 ret = ocfs2_rw_lock(inode, 0);
2341 /* communicate with ocfs2_dio_end_io */
2342 ocfs2_iocb_set_rw_locked(iocb, rw_level);
2346 * We're fine letting folks race truncates and extending
2347 * writes with read across the cluster, just like they can
2348 * locally. Hence no rw_lock during read.
2350 * Take and drop the meta data lock to update inode fields
2351 * like i_size. This allows the checks down below
2352 * generic_file_aio_read() a chance of actually working.
2354 ret = ocfs2_inode_lock_atime(inode, filp->f_vfsmnt, &lock_level);
2359 ocfs2_inode_unlock(inode, lock_level);
2361 ret = generic_file_aio_read(iocb, iov, nr_segs, iocb->ki_pos);
2363 mlog(0, "generic_file_aio_read returned -EINVAL\n");
2365 /* buffered aio wouldn't have proper lock coverage today */
2366 BUG_ON(ret == -EIOCBQUEUED && !(filp->f_flags & O_DIRECT));
2368 /* see ocfs2_file_aio_write */
2369 if (ret == -EIOCBQUEUED || !ocfs2_iocb_is_rw_locked(iocb)) {
2376 up_read(&inode->i_alloc_sem);
2378 ocfs2_rw_unlock(inode, rw_level);
2384 const struct inode_operations ocfs2_file_iops = {
2385 .setattr = ocfs2_setattr,
2386 .getattr = ocfs2_getattr,
2387 .permission = ocfs2_permission,
2388 .setxattr = generic_setxattr,
2389 .getxattr = generic_getxattr,
2390 .listxattr = ocfs2_listxattr,
2391 .removexattr = generic_removexattr,
2392 .fallocate = ocfs2_fallocate,
2393 .fiemap = ocfs2_fiemap,
2396 const struct inode_operations ocfs2_special_file_iops = {
2397 .setattr = ocfs2_setattr,
2398 .getattr = ocfs2_getattr,
2399 .permission = ocfs2_permission,
2403 * Other than ->lock, keep ocfs2_fops and ocfs2_dops in sync with
2404 * ocfs2_fops_no_plocks and ocfs2_dops_no_plocks!
2406 const struct file_operations ocfs2_fops = {
2407 .llseek = generic_file_llseek,
2408 .read = do_sync_read,
2409 .write = do_sync_write,
2411 .fsync = ocfs2_sync_file,
2412 .release = ocfs2_file_release,
2413 .open = ocfs2_file_open,
2414 .aio_read = ocfs2_file_aio_read,
2415 .aio_write = ocfs2_file_aio_write,
2416 .unlocked_ioctl = ocfs2_ioctl,
2417 #ifdef CONFIG_COMPAT
2418 .compat_ioctl = ocfs2_compat_ioctl,
2421 .flock = ocfs2_flock,
2422 .splice_read = ocfs2_file_splice_read,
2423 .splice_write = ocfs2_file_splice_write,
2426 const struct file_operations ocfs2_dops = {
2427 .llseek = generic_file_llseek,
2428 .read = generic_read_dir,
2429 .readdir = ocfs2_readdir,
2430 .fsync = ocfs2_sync_file,
2431 .release = ocfs2_dir_release,
2432 .open = ocfs2_dir_open,
2433 .unlocked_ioctl = ocfs2_ioctl,
2434 #ifdef CONFIG_COMPAT
2435 .compat_ioctl = ocfs2_compat_ioctl,
2438 .flock = ocfs2_flock,
2442 * POSIX-lockless variants of our file_operations.
2444 * These will be used if the underlying cluster stack does not support
2445 * posix file locking, if the user passes the "localflocks" mount
2446 * option, or if we have a local-only fs.
2448 * ocfs2_flock is in here because all stacks handle UNIX file locks,
2449 * so we still want it in the case of no stack support for
2450 * plocks. Internally, it will do the right thing when asked to ignore
2453 const struct file_operations ocfs2_fops_no_plocks = {
2454 .llseek = generic_file_llseek,
2455 .read = do_sync_read,
2456 .write = do_sync_write,
2458 .fsync = ocfs2_sync_file,
2459 .release = ocfs2_file_release,
2460 .open = ocfs2_file_open,
2461 .aio_read = ocfs2_file_aio_read,
2462 .aio_write = ocfs2_file_aio_write,
2463 .unlocked_ioctl = ocfs2_ioctl,
2464 #ifdef CONFIG_COMPAT
2465 .compat_ioctl = ocfs2_compat_ioctl,
2467 .flock = ocfs2_flock,
2468 .splice_read = ocfs2_file_splice_read,
2469 .splice_write = ocfs2_file_splice_write,
2472 const struct file_operations ocfs2_dops_no_plocks = {
2473 .llseek = generic_file_llseek,
2474 .read = generic_read_dir,
2475 .readdir = ocfs2_readdir,
2476 .fsync = ocfs2_sync_file,
2477 .release = ocfs2_dir_release,
2478 .open = ocfs2_dir_open,
2479 .unlocked_ioctl = ocfs2_ioctl,
2480 #ifdef CONFIG_COMPAT
2481 .compat_ioctl = ocfs2_compat_ioctl,
2483 .flock = ocfs2_flock,