xref: /linux/fs/ocfs2/file.c (revision f3d9478b2ce468c3115b02ecae7e975990697f15)
1 /* -*- mode: c; c-basic-offset: 8; -*-
2  * vim: noexpandtab sw=8 ts=8 sts=0:
3  *
4  * file.c
5  *
6  * File open, close, extend, truncate
7  *
8  * Copyright (C) 2002, 2004 Oracle.  All rights reserved.
9  *
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.
14  *
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.
19  *
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.
24  */
25 
26 #include <linux/capability.h>
27 #include <linux/fs.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 
34 #define MLOG_MASK_PREFIX ML_INODE
35 #include <cluster/masklog.h>
36 
37 #include "ocfs2.h"
38 
39 #include "alloc.h"
40 #include "aops.h"
41 #include "dir.h"
42 #include "dlmglue.h"
43 #include "extent_map.h"
44 #include "file.h"
45 #include "sysfile.h"
46 #include "inode.h"
47 #include "journal.h"
48 #include "mmap.h"
49 #include "suballoc.h"
50 #include "super.h"
51 
52 #include "buffer_head_io.h"
53 
54 static int ocfs2_sync_inode(struct inode *inode)
55 {
56 	filemap_fdatawrite(inode->i_mapping);
57 	return sync_mapping_buffers(inode->i_mapping);
58 }
59 
60 static int ocfs2_file_open(struct inode *inode, struct file *file)
61 {
62 	int status;
63 	int mode = file->f_flags;
64 	struct ocfs2_inode_info *oi = OCFS2_I(inode);
65 
66 	mlog_entry("(0x%p, 0x%p, '%.*s')\n", inode, file,
67 		   file->f_dentry->d_name.len, file->f_dentry->d_name.name);
68 
69 	spin_lock(&oi->ip_lock);
70 
71 	/* Check that the inode hasn't been wiped from disk by another
72 	 * node. If it hasn't then we're safe as long as we hold the
73 	 * spin lock until our increment of open count. */
74 	if (OCFS2_I(inode)->ip_flags & OCFS2_INODE_DELETED) {
75 		spin_unlock(&oi->ip_lock);
76 
77 		status = -ENOENT;
78 		goto leave;
79 	}
80 
81 	if (mode & O_DIRECT)
82 		oi->ip_flags |= OCFS2_INODE_OPEN_DIRECT;
83 
84 	oi->ip_open_count++;
85 	spin_unlock(&oi->ip_lock);
86 	status = 0;
87 leave:
88 	mlog_exit(status);
89 	return status;
90 }
91 
92 static int ocfs2_file_release(struct inode *inode, struct file *file)
93 {
94 	struct ocfs2_inode_info *oi = OCFS2_I(inode);
95 
96 	mlog_entry("(0x%p, 0x%p, '%.*s')\n", inode, file,
97 		       file->f_dentry->d_name.len,
98 		       file->f_dentry->d_name.name);
99 
100 	spin_lock(&oi->ip_lock);
101 	if (!--oi->ip_open_count)
102 		oi->ip_flags &= ~OCFS2_INODE_OPEN_DIRECT;
103 	spin_unlock(&oi->ip_lock);
104 
105 	mlog_exit(0);
106 
107 	return 0;
108 }
109 
110 static int ocfs2_sync_file(struct file *file,
111 			   struct dentry *dentry,
112 			   int datasync)
113 {
114 	int err = 0;
115 	journal_t *journal;
116 	struct inode *inode = dentry->d_inode;
117 	struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
118 
119 	mlog_entry("(0x%p, 0x%p, %d, '%.*s')\n", file, dentry, datasync,
120 		   dentry->d_name.len, dentry->d_name.name);
121 
122 	err = ocfs2_sync_inode(dentry->d_inode);
123 	if (err)
124 		goto bail;
125 
126 	journal = osb->journal->j_journal;
127 	err = journal_force_commit(journal);
128 
129 bail:
130 	mlog_exit(err);
131 
132 	return (err < 0) ? -EIO : 0;
133 }
134 
135 int ocfs2_set_inode_size(struct ocfs2_journal_handle *handle,
136 			 struct inode *inode,
137 			 struct buffer_head *fe_bh,
138 			 u64 new_i_size)
139 {
140 	int status;
141 
142 	mlog_entry_void();
143 	i_size_write(inode, new_i_size);
144 	inode->i_blocks = ocfs2_align_bytes_to_sectors(new_i_size);
145 	inode->i_ctime = inode->i_mtime = CURRENT_TIME;
146 
147 	status = ocfs2_mark_inode_dirty(handle, inode, fe_bh);
148 	if (status < 0) {
149 		mlog_errno(status);
150 		goto bail;
151 	}
152 
153 bail:
154 	mlog_exit(status);
155 	return status;
156 }
157 
158 static int ocfs2_simple_size_update(struct inode *inode,
159 				    struct buffer_head *di_bh,
160 				    u64 new_i_size)
161 {
162 	int ret;
163 	struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
164 	struct ocfs2_journal_handle *handle = NULL;
165 
166 	handle = ocfs2_start_trans(osb, NULL,
167 				   OCFS2_INODE_UPDATE_CREDITS);
168 	if (handle == NULL) {
169 		ret = -ENOMEM;
170 		mlog_errno(ret);
171 		goto out;
172 	}
173 
174 	ret = ocfs2_set_inode_size(handle, inode, di_bh,
175 				   new_i_size);
176 	if (ret < 0)
177 		mlog_errno(ret);
178 
179 	ocfs2_commit_trans(handle);
180 out:
181 	return ret;
182 }
183 
184 static int ocfs2_orphan_for_truncate(struct ocfs2_super *osb,
185 				     struct inode *inode,
186 				     struct buffer_head *fe_bh,
187 				     u64 new_i_size)
188 {
189 	int status;
190 	struct ocfs2_journal_handle *handle;
191 
192 	mlog_entry_void();
193 
194 	/* TODO: This needs to actually orphan the inode in this
195 	 * transaction. */
196 
197 	handle = ocfs2_start_trans(osb, NULL, OCFS2_INODE_UPDATE_CREDITS);
198 	if (IS_ERR(handle)) {
199 		status = PTR_ERR(handle);
200 		mlog_errno(status);
201 		goto out;
202 	}
203 
204 	status = ocfs2_set_inode_size(handle, inode, fe_bh, new_i_size);
205 	if (status < 0)
206 		mlog_errno(status);
207 
208 	ocfs2_commit_trans(handle);
209 out:
210 	mlog_exit(status);
211 	return status;
212 }
213 
214 static int ocfs2_truncate_file(struct inode *inode,
215 			       struct buffer_head *di_bh,
216 			       u64 new_i_size)
217 {
218 	int status = 0;
219 	struct ocfs2_dinode *fe = NULL;
220 	struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
221 	struct ocfs2_truncate_context *tc = NULL;
222 
223 	mlog_entry("(inode = %llu, new_i_size = %llu\n",
224 		   (unsigned long long)OCFS2_I(inode)->ip_blkno,
225 		   (unsigned long long)new_i_size);
226 
227 	truncate_inode_pages(inode->i_mapping, new_i_size);
228 
229 	fe = (struct ocfs2_dinode *) di_bh->b_data;
230 	if (!OCFS2_IS_VALID_DINODE(fe)) {
231 		OCFS2_RO_ON_INVALID_DINODE(inode->i_sb, fe);
232 		status = -EIO;
233 		goto bail;
234 	}
235 
236 	mlog_bug_on_msg(le64_to_cpu(fe->i_size) != i_size_read(inode),
237 			"Inode %llu, inode i_size = %lld != di "
238 			"i_size = %llu, i_flags = 0x%x\n",
239 			(unsigned long long)OCFS2_I(inode)->ip_blkno,
240 			i_size_read(inode),
241 			(unsigned long long)le64_to_cpu(fe->i_size),
242 			le32_to_cpu(fe->i_flags));
243 
244 	if (new_i_size > le64_to_cpu(fe->i_size)) {
245 		mlog(0, "asked to truncate file with size (%llu) to size (%llu)!\n",
246 		     (unsigned long long)le64_to_cpu(fe->i_size),
247 		     (unsigned long long)new_i_size);
248 		status = -EINVAL;
249 		mlog_errno(status);
250 		goto bail;
251 	}
252 
253 	mlog(0, "inode %llu, i_size = %llu, new_i_size = %llu\n",
254 	     (unsigned long long)le64_to_cpu(fe->i_blkno),
255 	     (unsigned long long)le64_to_cpu(fe->i_size),
256 	     (unsigned long long)new_i_size);
257 
258 	/* lets handle the simple truncate cases before doing any more
259 	 * cluster locking. */
260 	if (new_i_size == le64_to_cpu(fe->i_size))
261 		goto bail;
262 
263 	/* This forces other nodes to sync and drop their pages. Do
264 	 * this even if we have a truncate without allocation change -
265 	 * ocfs2 cluster sizes can be much greater than page size, so
266 	 * we have to truncate them anyway.  */
267 	status = ocfs2_data_lock(inode, 1);
268 	if (status < 0) {
269 		mlog_errno(status);
270 		goto bail;
271 	}
272 	ocfs2_data_unlock(inode, 1);
273 
274 	if (le32_to_cpu(fe->i_clusters) ==
275 	    ocfs2_clusters_for_bytes(osb->sb, new_i_size)) {
276 		mlog(0, "fe->i_clusters = %u, so we do a simple truncate\n",
277 		     fe->i_clusters);
278 		/* No allocation change is required, so lets fast path
279 		 * this truncate. */
280 		status = ocfs2_simple_size_update(inode, di_bh, new_i_size);
281 		if (status < 0)
282 			mlog_errno(status);
283 		goto bail;
284 	}
285 
286 	/* alright, we're going to need to do a full blown alloc size
287 	 * change. Orphan the inode so that recovery can complete the
288 	 * truncate if necessary. This does the task of marking
289 	 * i_size. */
290 	status = ocfs2_orphan_for_truncate(osb, inode, di_bh, new_i_size);
291 	if (status < 0) {
292 		mlog_errno(status);
293 		goto bail;
294 	}
295 
296 	status = ocfs2_prepare_truncate(osb, inode, di_bh, &tc);
297 	if (status < 0) {
298 		mlog_errno(status);
299 		goto bail;
300 	}
301 
302 	status = ocfs2_commit_truncate(osb, inode, di_bh, tc);
303 	if (status < 0) {
304 		mlog_errno(status);
305 		goto bail;
306 	}
307 
308 	/* TODO: orphan dir cleanup here. */
309 bail:
310 
311 	mlog_exit(status);
312 	return status;
313 }
314 
315 /*
316  * extend allocation only here.
317  * we'll update all the disk stuff, and oip->alloc_size
318  *
319  * expect stuff to be locked, a transaction started and enough data /
320  * metadata reservations in the contexts.
321  *
322  * Will return -EAGAIN, and a reason if a restart is needed.
323  * If passed in, *reason will always be set, even in error.
324  */
325 int ocfs2_do_extend_allocation(struct ocfs2_super *osb,
326 			       struct inode *inode,
327 			       u32 clusters_to_add,
328 			       struct buffer_head *fe_bh,
329 			       struct ocfs2_journal_handle *handle,
330 			       struct ocfs2_alloc_context *data_ac,
331 			       struct ocfs2_alloc_context *meta_ac,
332 			       enum ocfs2_alloc_restarted *reason_ret)
333 {
334 	int status = 0;
335 	int free_extents;
336 	struct ocfs2_dinode *fe = (struct ocfs2_dinode *) fe_bh->b_data;
337 	enum ocfs2_alloc_restarted reason = RESTART_NONE;
338 	u32 bit_off, num_bits;
339 	u64 block;
340 
341 	BUG_ON(!clusters_to_add);
342 
343 	free_extents = ocfs2_num_free_extents(osb, inode, fe);
344 	if (free_extents < 0) {
345 		status = free_extents;
346 		mlog_errno(status);
347 		goto leave;
348 	}
349 
350 	/* there are two cases which could cause us to EAGAIN in the
351 	 * we-need-more-metadata case:
352 	 * 1) we haven't reserved *any*
353 	 * 2) we are so fragmented, we've needed to add metadata too
354 	 *    many times. */
355 	if (!free_extents && !meta_ac) {
356 		mlog(0, "we haven't reserved any metadata!\n");
357 		status = -EAGAIN;
358 		reason = RESTART_META;
359 		goto leave;
360 	} else if ((!free_extents)
361 		   && (ocfs2_alloc_context_bits_left(meta_ac)
362 		       < ocfs2_extend_meta_needed(fe))) {
363 		mlog(0, "filesystem is really fragmented...\n");
364 		status = -EAGAIN;
365 		reason = RESTART_META;
366 		goto leave;
367 	}
368 
369 	status = ocfs2_claim_clusters(osb, handle, data_ac, 1,
370 				      &bit_off, &num_bits);
371 	if (status < 0) {
372 		if (status != -ENOSPC)
373 			mlog_errno(status);
374 		goto leave;
375 	}
376 
377 	BUG_ON(num_bits > clusters_to_add);
378 
379 	/* reserve our write early -- insert_extent may update the inode */
380 	status = ocfs2_journal_access(handle, inode, fe_bh,
381 				      OCFS2_JOURNAL_ACCESS_WRITE);
382 	if (status < 0) {
383 		mlog_errno(status);
384 		goto leave;
385 	}
386 
387 	block = ocfs2_clusters_to_blocks(osb->sb, bit_off);
388 	mlog(0, "Allocating %u clusters at block %u for inode %llu\n",
389 	     num_bits, bit_off, (unsigned long long)OCFS2_I(inode)->ip_blkno);
390 	status = ocfs2_insert_extent(osb, handle, inode, fe_bh, block,
391 				     num_bits, meta_ac);
392 	if (status < 0) {
393 		mlog_errno(status);
394 		goto leave;
395 	}
396 
397 	le32_add_cpu(&fe->i_clusters, num_bits);
398 	spin_lock(&OCFS2_I(inode)->ip_lock);
399 	OCFS2_I(inode)->ip_clusters = le32_to_cpu(fe->i_clusters);
400 	spin_unlock(&OCFS2_I(inode)->ip_lock);
401 
402 	status = ocfs2_journal_dirty(handle, fe_bh);
403 	if (status < 0) {
404 		mlog_errno(status);
405 		goto leave;
406 	}
407 
408 	clusters_to_add -= num_bits;
409 
410 	if (clusters_to_add) {
411 		mlog(0, "need to alloc once more, clusters = %u, wanted = "
412 		     "%u\n", fe->i_clusters, clusters_to_add);
413 		status = -EAGAIN;
414 		reason = RESTART_TRANS;
415 	}
416 
417 leave:
418 	mlog_exit(status);
419 	if (reason_ret)
420 		*reason_ret = reason;
421 	return status;
422 }
423 
424 static int ocfs2_extend_allocation(struct inode *inode,
425 				   u32 clusters_to_add)
426 {
427 	int status = 0;
428 	int restart_func = 0;
429 	int drop_alloc_sem = 0;
430 	int credits, num_free_extents;
431 	u32 prev_clusters;
432 	struct buffer_head *bh = NULL;
433 	struct ocfs2_dinode *fe = NULL;
434 	struct ocfs2_journal_handle *handle = NULL;
435 	struct ocfs2_alloc_context *data_ac = NULL;
436 	struct ocfs2_alloc_context *meta_ac = NULL;
437 	enum ocfs2_alloc_restarted why;
438 	struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
439 
440 	mlog_entry("(clusters_to_add = %u)\n", clusters_to_add);
441 
442 	status = ocfs2_read_block(osb, OCFS2_I(inode)->ip_blkno, &bh,
443 				  OCFS2_BH_CACHED, inode);
444 	if (status < 0) {
445 		mlog_errno(status);
446 		goto leave;
447 	}
448 
449 	fe = (struct ocfs2_dinode *) bh->b_data;
450 	if (!OCFS2_IS_VALID_DINODE(fe)) {
451 		OCFS2_RO_ON_INVALID_DINODE(inode->i_sb, fe);
452 		status = -EIO;
453 		goto leave;
454 	}
455 
456 restart_all:
457 	BUG_ON(le32_to_cpu(fe->i_clusters) != OCFS2_I(inode)->ip_clusters);
458 
459 	mlog(0, "extend inode %llu, i_size = %lld, fe->i_clusters = %u, "
460 	     "clusters_to_add = %u\n",
461 	     (unsigned long long)OCFS2_I(inode)->ip_blkno, i_size_read(inode),
462 	     fe->i_clusters, clusters_to_add);
463 
464 	handle = ocfs2_alloc_handle(osb);
465 	if (handle == NULL) {
466 		status = -ENOMEM;
467 		mlog_errno(status);
468 		goto leave;
469 	}
470 
471 	num_free_extents = ocfs2_num_free_extents(osb,
472 						  inode,
473 						  fe);
474 	if (num_free_extents < 0) {
475 		status = num_free_extents;
476 		mlog_errno(status);
477 		goto leave;
478 	}
479 
480 	if (!num_free_extents) {
481 		status = ocfs2_reserve_new_metadata(osb,
482 						    handle,
483 						    fe,
484 						    &meta_ac);
485 		if (status < 0) {
486 			if (status != -ENOSPC)
487 				mlog_errno(status);
488 			goto leave;
489 		}
490 	}
491 
492 	status = ocfs2_reserve_clusters(osb,
493 					handle,
494 					clusters_to_add,
495 					&data_ac);
496 	if (status < 0) {
497 		if (status != -ENOSPC)
498 			mlog_errno(status);
499 		goto leave;
500 	}
501 
502 	/* blocks peope in read/write from reading our allocation
503 	 * until we're done changing it. We depend on i_mutex to block
504 	 * other extend/truncate calls while we're here. Ordering wrt
505 	 * start_trans is important here -- always do it before! */
506 	down_write(&OCFS2_I(inode)->ip_alloc_sem);
507 	drop_alloc_sem = 1;
508 
509 	credits = ocfs2_calc_extend_credits(osb->sb, fe, clusters_to_add);
510 	handle = ocfs2_start_trans(osb, handle, credits);
511 	if (IS_ERR(handle)) {
512 		status = PTR_ERR(handle);
513 		handle = NULL;
514 		mlog_errno(status);
515 		goto leave;
516 	}
517 
518 restarted_transaction:
519 	/* reserve a write to the file entry early on - that we if we
520 	 * run out of credits in the allocation path, we can still
521 	 * update i_size. */
522 	status = ocfs2_journal_access(handle, inode, bh,
523 				      OCFS2_JOURNAL_ACCESS_WRITE);
524 	if (status < 0) {
525 		mlog_errno(status);
526 		goto leave;
527 	}
528 
529 	prev_clusters = OCFS2_I(inode)->ip_clusters;
530 
531 	status = ocfs2_do_extend_allocation(osb,
532 					    inode,
533 					    clusters_to_add,
534 					    bh,
535 					    handle,
536 					    data_ac,
537 					    meta_ac,
538 					    &why);
539 	if ((status < 0) && (status != -EAGAIN)) {
540 		if (status != -ENOSPC)
541 			mlog_errno(status);
542 		goto leave;
543 	}
544 
545 	status = ocfs2_journal_dirty(handle, bh);
546 	if (status < 0) {
547 		mlog_errno(status);
548 		goto leave;
549 	}
550 
551 	spin_lock(&OCFS2_I(inode)->ip_lock);
552 	clusters_to_add -= (OCFS2_I(inode)->ip_clusters - prev_clusters);
553 	spin_unlock(&OCFS2_I(inode)->ip_lock);
554 
555 	if (why != RESTART_NONE && clusters_to_add) {
556 		if (why == RESTART_META) {
557 			mlog(0, "restarting function.\n");
558 			restart_func = 1;
559 		} else {
560 			BUG_ON(why != RESTART_TRANS);
561 
562 			mlog(0, "restarting transaction.\n");
563 			/* TODO: This can be more intelligent. */
564 			credits = ocfs2_calc_extend_credits(osb->sb,
565 							    fe,
566 							    clusters_to_add);
567 			status = ocfs2_extend_trans(handle, credits);
568 			if (status < 0) {
569 				/* handle still has to be committed at
570 				 * this point. */
571 				status = -ENOMEM;
572 				mlog_errno(status);
573 				goto leave;
574 			}
575 			goto restarted_transaction;
576 		}
577 	}
578 
579 	mlog(0, "fe: i_clusters = %u, i_size=%llu\n",
580 	     fe->i_clusters, (unsigned long long)fe->i_size);
581 	mlog(0, "inode: ip_clusters=%u, i_size=%lld\n",
582 	     OCFS2_I(inode)->ip_clusters, i_size_read(inode));
583 
584 leave:
585 	if (drop_alloc_sem) {
586 		up_write(&OCFS2_I(inode)->ip_alloc_sem);
587 		drop_alloc_sem = 0;
588 	}
589 	if (handle) {
590 		ocfs2_commit_trans(handle);
591 		handle = NULL;
592 	}
593 	if (data_ac) {
594 		ocfs2_free_alloc_context(data_ac);
595 		data_ac = NULL;
596 	}
597 	if (meta_ac) {
598 		ocfs2_free_alloc_context(meta_ac);
599 		meta_ac = NULL;
600 	}
601 	if ((!status) && restart_func) {
602 		restart_func = 0;
603 		goto restart_all;
604 	}
605 	if (bh) {
606 		brelse(bh);
607 		bh = NULL;
608 	}
609 
610 	mlog_exit(status);
611 	return status;
612 }
613 
614 /* Some parts of this taken from generic_cont_expand, which turned out
615  * to be too fragile to do exactly what we need without us having to
616  * worry about recursive locking in ->prepare_write() and
617  * ->commit_write(). */
618 static int ocfs2_write_zero_page(struct inode *inode,
619 				 u64 size)
620 {
621 	struct address_space *mapping = inode->i_mapping;
622 	struct page *page;
623 	unsigned long index;
624 	unsigned int offset;
625 	struct ocfs2_journal_handle *handle = NULL;
626 	int ret;
627 
628 	offset = (size & (PAGE_CACHE_SIZE-1)); /* Within page */
629 	/* ugh.  in prepare/commit_write, if from==to==start of block, we
630 	** skip the prepare.  make sure we never send an offset for the start
631 	** of a block
632 	*/
633 	if ((offset & (inode->i_sb->s_blocksize - 1)) == 0) {
634 		offset++;
635 	}
636 	index = size >> PAGE_CACHE_SHIFT;
637 
638 	page = grab_cache_page(mapping, index);
639 	if (!page) {
640 		ret = -ENOMEM;
641 		mlog_errno(ret);
642 		goto out;
643 	}
644 
645 	ret = ocfs2_prepare_write_nolock(inode, page, offset, offset);
646 	if (ret < 0) {
647 		mlog_errno(ret);
648 		goto out_unlock;
649 	}
650 
651 	if (ocfs2_should_order_data(inode)) {
652 		handle = ocfs2_start_walk_page_trans(inode, page, offset,
653 						     offset);
654 		if (IS_ERR(handle)) {
655 			ret = PTR_ERR(handle);
656 			handle = NULL;
657 			goto out_unlock;
658 		}
659 	}
660 
661 	/* must not update i_size! */
662 	ret = block_commit_write(page, offset, offset);
663 	if (ret < 0)
664 		mlog_errno(ret);
665 	else
666 		ret = 0;
667 
668 	if (handle)
669 		ocfs2_commit_trans(handle);
670 out_unlock:
671 	unlock_page(page);
672 	page_cache_release(page);
673 out:
674 	return ret;
675 }
676 
677 static int ocfs2_zero_extend(struct inode *inode,
678 			     u64 zero_to_size)
679 {
680 	int ret = 0;
681 	u64 start_off;
682 	struct super_block *sb = inode->i_sb;
683 
684 	start_off = ocfs2_align_bytes_to_blocks(sb, i_size_read(inode));
685 	while (start_off < zero_to_size) {
686 		ret = ocfs2_write_zero_page(inode, start_off);
687 		if (ret < 0) {
688 			mlog_errno(ret);
689 			goto out;
690 		}
691 
692 		start_off += sb->s_blocksize;
693 	}
694 
695 out:
696 	return ret;
697 }
698 
699 /*
700  * A tail_to_skip value > 0 indicates that we're being called from
701  * ocfs2_file_aio_write(). This has the following implications:
702  *
703  * - we don't want to update i_size
704  * - di_bh will be NULL, which is fine because it's only used in the
705  *   case where we want to update i_size.
706  * - ocfs2_zero_extend() will then only be filling the hole created
707  *   between i_size and the start of the write.
708  */
709 static int ocfs2_extend_file(struct inode *inode,
710 			     struct buffer_head *di_bh,
711 			     u64 new_i_size,
712 			     size_t tail_to_skip)
713 {
714 	int ret = 0;
715 	u32 clusters_to_add;
716 
717 	BUG_ON(!tail_to_skip && !di_bh);
718 
719 	/* setattr sometimes calls us like this. */
720 	if (new_i_size == 0)
721 		goto out;
722 
723 	if (i_size_read(inode) == new_i_size)
724   		goto out;
725 	BUG_ON(new_i_size < i_size_read(inode));
726 
727 	clusters_to_add = ocfs2_clusters_for_bytes(inode->i_sb, new_i_size) -
728 		OCFS2_I(inode)->ip_clusters;
729 
730 	if (clusters_to_add) {
731 		/*
732 		 * protect the pages that ocfs2_zero_extend is going to
733 		 * be pulling into the page cache.. we do this before the
734 		 * metadata extend so that we don't get into the situation
735 		 * where we've extended the metadata but can't get the data
736 		 * lock to zero.
737 		 */
738 		ret = ocfs2_data_lock(inode, 1);
739 		if (ret < 0) {
740 			mlog_errno(ret);
741 			goto out;
742 		}
743 
744 		ret = ocfs2_extend_allocation(inode, clusters_to_add);
745 		if (ret < 0) {
746 			mlog_errno(ret);
747 			goto out_unlock;
748 		}
749 
750 		ret = ocfs2_zero_extend(inode, (u64)new_i_size - tail_to_skip);
751 		if (ret < 0) {
752 			mlog_errno(ret);
753 			goto out_unlock;
754 		}
755 	}
756 
757 	if (!tail_to_skip) {
758 		/* We're being called from ocfs2_setattr() which wants
759 		 * us to update i_size */
760 		ret = ocfs2_simple_size_update(inode, di_bh, new_i_size);
761 		if (ret < 0)
762 			mlog_errno(ret);
763 	}
764 
765 out_unlock:
766 	if (clusters_to_add) /* this is the only case in which we lock */
767 		ocfs2_data_unlock(inode, 1);
768 
769 out:
770 	return ret;
771 }
772 
773 int ocfs2_setattr(struct dentry *dentry, struct iattr *attr)
774 {
775 	int status = 0, size_change;
776 	struct inode *inode = dentry->d_inode;
777 	struct super_block *sb = inode->i_sb;
778 	struct ocfs2_super *osb = OCFS2_SB(sb);
779 	struct buffer_head *bh = NULL;
780 	struct ocfs2_journal_handle *handle = NULL;
781 
782 	mlog_entry("(0x%p, '%.*s')\n", dentry,
783 	           dentry->d_name.len, dentry->d_name.name);
784 
785 	if (attr->ia_valid & ATTR_MODE)
786 		mlog(0, "mode change: %d\n", attr->ia_mode);
787 	if (attr->ia_valid & ATTR_UID)
788 		mlog(0, "uid change: %d\n", attr->ia_uid);
789 	if (attr->ia_valid & ATTR_GID)
790 		mlog(0, "gid change: %d\n", attr->ia_gid);
791 	if (attr->ia_valid & ATTR_SIZE)
792 		mlog(0, "size change...\n");
793 	if (attr->ia_valid & (ATTR_ATIME | ATTR_MTIME | ATTR_CTIME))
794 		mlog(0, "time change...\n");
795 
796 #define OCFS2_VALID_ATTRS (ATTR_ATIME | ATTR_MTIME | ATTR_CTIME | ATTR_SIZE \
797 			   | ATTR_GID | ATTR_UID | ATTR_MODE)
798 	if (!(attr->ia_valid & OCFS2_VALID_ATTRS)) {
799 		mlog(0, "can't handle attrs: 0x%x\n", attr->ia_valid);
800 		return 0;
801 	}
802 
803 	status = inode_change_ok(inode, attr);
804 	if (status)
805 		return status;
806 
807 	size_change = S_ISREG(inode->i_mode) && attr->ia_valid & ATTR_SIZE;
808 	if (size_change) {
809 		status = ocfs2_rw_lock(inode, 1);
810 		if (status < 0) {
811 			mlog_errno(status);
812 			goto bail;
813 		}
814 	}
815 
816 	status = ocfs2_meta_lock(inode, NULL, &bh, 1);
817 	if (status < 0) {
818 		if (status != -ENOENT)
819 			mlog_errno(status);
820 		goto bail_unlock_rw;
821 	}
822 
823 	if (size_change && attr->ia_size != i_size_read(inode)) {
824 		if (i_size_read(inode) > attr->ia_size)
825 			status = ocfs2_truncate_file(inode, bh, attr->ia_size);
826 		else
827 			status = ocfs2_extend_file(inode, bh, attr->ia_size, 0);
828 		if (status < 0) {
829 			if (status != -ENOSPC)
830 				mlog_errno(status);
831 			status = -ENOSPC;
832 			goto bail_unlock;
833 		}
834 	}
835 
836 	handle = ocfs2_start_trans(osb, NULL, OCFS2_INODE_UPDATE_CREDITS);
837 	if (IS_ERR(handle)) {
838 		status = PTR_ERR(handle);
839 		mlog_errno(status);
840 		goto bail_unlock;
841 	}
842 
843 	status = inode_setattr(inode, attr);
844 	if (status < 0) {
845 		mlog_errno(status);
846 		goto bail_commit;
847 	}
848 
849 	status = ocfs2_mark_inode_dirty(handle, inode, bh);
850 	if (status < 0)
851 		mlog_errno(status);
852 
853 bail_commit:
854 	ocfs2_commit_trans(handle);
855 bail_unlock:
856 	ocfs2_meta_unlock(inode, 1);
857 bail_unlock_rw:
858 	if (size_change)
859 		ocfs2_rw_unlock(inode, 1);
860 bail:
861 	if (bh)
862 		brelse(bh);
863 
864 	mlog_exit(status);
865 	return status;
866 }
867 
868 int ocfs2_getattr(struct vfsmount *mnt,
869 		  struct dentry *dentry,
870 		  struct kstat *stat)
871 {
872 	struct inode *inode = dentry->d_inode;
873 	struct super_block *sb = dentry->d_inode->i_sb;
874 	struct ocfs2_super *osb = sb->s_fs_info;
875 	int err;
876 
877 	mlog_entry_void();
878 
879 	err = ocfs2_inode_revalidate(dentry);
880 	if (err) {
881 		if (err != -ENOENT)
882 			mlog_errno(err);
883 		goto bail;
884 	}
885 
886 	generic_fillattr(inode, stat);
887 
888 	/* We set the blksize from the cluster size for performance */
889 	stat->blksize = osb->s_clustersize;
890 
891 bail:
892 	mlog_exit(err);
893 
894 	return err;
895 }
896 
897 static int ocfs2_write_remove_suid(struct inode *inode)
898 {
899 	int ret;
900 	struct buffer_head *bh = NULL;
901 	struct ocfs2_inode_info *oi = OCFS2_I(inode);
902 	struct ocfs2_journal_handle *handle;
903 	struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
904 	struct ocfs2_dinode *di;
905 
906 	mlog_entry("(Inode %llu, mode 0%o)\n",
907 		   (unsigned long long)oi->ip_blkno, inode->i_mode);
908 
909 	handle = ocfs2_start_trans(osb, NULL, OCFS2_INODE_UPDATE_CREDITS);
910 	if (handle == NULL) {
911 		ret = -ENOMEM;
912 		mlog_errno(ret);
913 		goto out;
914 	}
915 
916 	ret = ocfs2_read_block(osb, oi->ip_blkno, &bh, OCFS2_BH_CACHED, inode);
917 	if (ret < 0) {
918 		mlog_errno(ret);
919 		goto out_trans;
920 	}
921 
922 	ret = ocfs2_journal_access(handle, inode, bh,
923 				   OCFS2_JOURNAL_ACCESS_WRITE);
924 	if (ret < 0) {
925 		mlog_errno(ret);
926 		goto out_bh;
927 	}
928 
929 	inode->i_mode &= ~S_ISUID;
930 	if ((inode->i_mode & S_ISGID) && (inode->i_mode & S_IXGRP))
931 		inode->i_mode &= ~S_ISGID;
932 
933 	di = (struct ocfs2_dinode *) bh->b_data;
934 	di->i_mode = cpu_to_le16(inode->i_mode);
935 
936 	ret = ocfs2_journal_dirty(handle, bh);
937 	if (ret < 0)
938 		mlog_errno(ret);
939 out_bh:
940 	brelse(bh);
941 out_trans:
942 	ocfs2_commit_trans(handle);
943 out:
944 	mlog_exit(ret);
945 	return ret;
946 }
947 
948 static inline int ocfs2_write_should_remove_suid(struct inode *inode)
949 {
950 	mode_t mode = inode->i_mode;
951 
952 	if (!capable(CAP_FSETID)) {
953 		if (unlikely(mode & S_ISUID))
954 			return 1;
955 
956 		if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
957 			return 1;
958 	}
959 	return 0;
960 }
961 
962 static ssize_t ocfs2_file_aio_write(struct kiocb *iocb,
963 				    const char __user *buf,
964 				    size_t count,
965 				    loff_t pos)
966 {
967 	struct iovec local_iov = { .iov_base = (void __user *)buf,
968 				   .iov_len = count };
969 	int ret, rw_level = -1, meta_level = -1, have_alloc_sem = 0;
970 	u32 clusters;
971 	struct file *filp = iocb->ki_filp;
972 	struct inode *inode = filp->f_dentry->d_inode;
973 	loff_t newsize, saved_pos;
974 
975 	mlog_entry("(0x%p, 0x%p, %u, '%.*s')\n", filp, buf,
976 		   (unsigned int)count,
977 		   filp->f_dentry->d_name.len,
978 		   filp->f_dentry->d_name.name);
979 
980 	/* happy write of zero bytes */
981 	if (count == 0)
982 		return 0;
983 
984 	if (!inode) {
985 		mlog(0, "bad inode\n");
986 		return -EIO;
987 	}
988 
989 	mutex_lock(&inode->i_mutex);
990 	/* to match setattr's i_mutex -> i_alloc_sem -> rw_lock ordering */
991 	if (filp->f_flags & O_DIRECT) {
992 		have_alloc_sem = 1;
993 		down_read(&inode->i_alloc_sem);
994 	}
995 
996 	/* concurrent O_DIRECT writes are allowed */
997 	rw_level = (filp->f_flags & O_DIRECT) ? 0 : 1;
998 	ret = ocfs2_rw_lock(inode, rw_level);
999 	if (ret < 0) {
1000 		rw_level = -1;
1001 		mlog_errno(ret);
1002 		goto out;
1003 	}
1004 
1005 	/*
1006 	 * We sample i_size under a read level meta lock to see if our write
1007 	 * is extending the file, if it is we back off and get a write level
1008 	 * meta lock.
1009 	 */
1010 	meta_level = (filp->f_flags & O_APPEND) ? 1 : 0;
1011 	for(;;) {
1012 		ret = ocfs2_meta_lock(inode, NULL, NULL, meta_level);
1013 		if (ret < 0) {
1014 			meta_level = -1;
1015 			mlog_errno(ret);
1016 			goto out;
1017 		}
1018 
1019 		/* Clear suid / sgid if necessary. We do this here
1020 		 * instead of later in the write path because
1021 		 * remove_suid() calls ->setattr without any hint that
1022 		 * we may have already done our cluster locking. Since
1023 		 * ocfs2_setattr() *must* take cluster locks to
1024 		 * proceeed, this will lead us to recursively lock the
1025 		 * inode. There's also the dinode i_size state which
1026 		 * can be lost via setattr during extending writes (we
1027 		 * set inode->i_size at the end of a write. */
1028 		if (ocfs2_write_should_remove_suid(inode)) {
1029 			if (meta_level == 0) {
1030 				ocfs2_meta_unlock(inode, meta_level);
1031 				meta_level = 1;
1032 				continue;
1033 			}
1034 
1035 			ret = ocfs2_write_remove_suid(inode);
1036 			if (ret < 0) {
1037 				mlog_errno(ret);
1038 				goto out;
1039 			}
1040 		}
1041 
1042 		/* work on a copy of ppos until we're sure that we won't have
1043 		 * to recalculate it due to relocking. */
1044 		if (filp->f_flags & O_APPEND) {
1045 			saved_pos = i_size_read(inode);
1046 			mlog(0, "O_APPEND: inode->i_size=%llu\n", saved_pos);
1047 		} else {
1048 			saved_pos = iocb->ki_pos;
1049 		}
1050 		newsize = count + saved_pos;
1051 
1052 		mlog(0, "pos=%lld newsize=%lld cursize=%lld\n",
1053 		     (long long) saved_pos, (long long) newsize,
1054 		     (long long) i_size_read(inode));
1055 
1056 		/* No need for a higher level metadata lock if we're
1057 		 * never going past i_size. */
1058 		if (newsize <= i_size_read(inode))
1059 			break;
1060 
1061 		if (meta_level == 0) {
1062 			ocfs2_meta_unlock(inode, meta_level);
1063 			meta_level = 1;
1064 			continue;
1065 		}
1066 
1067 		spin_lock(&OCFS2_I(inode)->ip_lock);
1068 		clusters = ocfs2_clusters_for_bytes(inode->i_sb, newsize) -
1069 			OCFS2_I(inode)->ip_clusters;
1070 		spin_unlock(&OCFS2_I(inode)->ip_lock);
1071 
1072 		mlog(0, "Writing at EOF, may need more allocation: "
1073 		     "i_size = %lld, newsize = %lld, need %u clusters\n",
1074 		     (long long) i_size_read(inode), (long long) newsize,
1075 		     clusters);
1076 
1077 		/* We only want to continue the rest of this loop if
1078 		 * our extend will actually require more
1079 		 * allocation. */
1080 		if (!clusters)
1081 			break;
1082 
1083 		ret = ocfs2_extend_file(inode, NULL, newsize, count);
1084 		if (ret < 0) {
1085 			if (ret != -ENOSPC)
1086 				mlog_errno(ret);
1087 			goto out;
1088 		}
1089 		break;
1090 	}
1091 
1092 	/* ok, we're done with i_size and alloc work */
1093 	iocb->ki_pos = saved_pos;
1094 	ocfs2_meta_unlock(inode, meta_level);
1095 	meta_level = -1;
1096 
1097 	/* communicate with ocfs2_dio_end_io */
1098 	ocfs2_iocb_set_rw_locked(iocb);
1099 
1100 	ret = generic_file_aio_write_nolock(iocb, &local_iov, 1, &iocb->ki_pos);
1101 
1102 	/* buffered aio wouldn't have proper lock coverage today */
1103 	BUG_ON(ret == -EIOCBQUEUED && !(filp->f_flags & O_DIRECT));
1104 
1105 	/*
1106 	 * deep in g_f_a_w_n()->ocfs2_direct_IO we pass in a ocfs2_dio_end_io
1107 	 * function pointer which is called when o_direct io completes so that
1108 	 * it can unlock our rw lock.  (it's the clustered equivalent of
1109 	 * i_alloc_sem; protects truncate from racing with pending ios).
1110 	 * Unfortunately there are error cases which call end_io and others
1111 	 * that don't.  so we don't have to unlock the rw_lock if either an
1112 	 * async dio is going to do it in the future or an end_io after an
1113 	 * error has already done it.
1114 	 */
1115 	if (ret == -EIOCBQUEUED || !ocfs2_iocb_is_rw_locked(iocb)) {
1116 		rw_level = -1;
1117 		have_alloc_sem = 0;
1118 	}
1119 
1120 out:
1121 	if (meta_level != -1)
1122 		ocfs2_meta_unlock(inode, meta_level);
1123 	if (have_alloc_sem)
1124 		up_read(&inode->i_alloc_sem);
1125 	if (rw_level != -1)
1126 		ocfs2_rw_unlock(inode, rw_level);
1127 	mutex_unlock(&inode->i_mutex);
1128 
1129 	mlog_exit(ret);
1130 	return ret;
1131 }
1132 
1133 static ssize_t ocfs2_file_aio_read(struct kiocb *iocb,
1134 				   char __user *buf,
1135 				   size_t count,
1136 				   loff_t pos)
1137 {
1138 	int ret = 0, rw_level = -1, have_alloc_sem = 0;
1139 	struct file *filp = iocb->ki_filp;
1140 	struct inode *inode = filp->f_dentry->d_inode;
1141 
1142 	mlog_entry("(0x%p, 0x%p, %u, '%.*s')\n", filp, buf,
1143 		   (unsigned int)count,
1144 		   filp->f_dentry->d_name.len,
1145 		   filp->f_dentry->d_name.name);
1146 
1147 	if (!inode) {
1148 		ret = -EINVAL;
1149 		mlog_errno(ret);
1150 		goto bail;
1151 	}
1152 
1153 	/*
1154 	 * buffered reads protect themselves in ->readpage().  O_DIRECT reads
1155 	 * need locks to protect pending reads from racing with truncate.
1156 	 */
1157 	if (filp->f_flags & O_DIRECT) {
1158 		down_read(&inode->i_alloc_sem);
1159 		have_alloc_sem = 1;
1160 
1161 		ret = ocfs2_rw_lock(inode, 0);
1162 		if (ret < 0) {
1163 			mlog_errno(ret);
1164 			goto bail;
1165 		}
1166 		rw_level = 0;
1167 		/* communicate with ocfs2_dio_end_io */
1168 		ocfs2_iocb_set_rw_locked(iocb);
1169 	}
1170 
1171 	/*
1172 	 * We're fine letting folks race truncates and extending
1173 	 * writes with read across the cluster, just like they can
1174 	 * locally. Hence no rw_lock during read.
1175 	 *
1176 	 * Take and drop the meta data lock to update inode fields
1177 	 * like i_size. This allows the checks down below
1178 	 * generic_file_aio_read() a chance of actually working.
1179 	 */
1180 	ret = ocfs2_meta_lock(inode, NULL, NULL, 0);
1181 	if (ret < 0) {
1182 		mlog_errno(ret);
1183 		goto bail;
1184 	}
1185 	ocfs2_meta_unlock(inode, 0);
1186 
1187 	ret = generic_file_aio_read(iocb, buf, count, iocb->ki_pos);
1188 	if (ret == -EINVAL)
1189 		mlog(ML_ERROR, "generic_file_aio_read returned -EINVAL\n");
1190 
1191 	/* buffered aio wouldn't have proper lock coverage today */
1192 	BUG_ON(ret == -EIOCBQUEUED && !(filp->f_flags & O_DIRECT));
1193 
1194 	/* see ocfs2_file_aio_write */
1195 	if (ret == -EIOCBQUEUED || !ocfs2_iocb_is_rw_locked(iocb)) {
1196 		rw_level = -1;
1197 		have_alloc_sem = 0;
1198 	}
1199 
1200 bail:
1201 	if (have_alloc_sem)
1202 		up_read(&inode->i_alloc_sem);
1203 	if (rw_level != -1)
1204 		ocfs2_rw_unlock(inode, rw_level);
1205 	mlog_exit(ret);
1206 
1207 	return ret;
1208 }
1209 
1210 struct inode_operations ocfs2_file_iops = {
1211 	.setattr	= ocfs2_setattr,
1212 	.getattr	= ocfs2_getattr,
1213 };
1214 
1215 struct inode_operations ocfs2_special_file_iops = {
1216 	.setattr	= ocfs2_setattr,
1217 	.getattr	= ocfs2_getattr,
1218 };
1219 
1220 const struct file_operations ocfs2_fops = {
1221 	.read		= do_sync_read,
1222 	.write		= do_sync_write,
1223 	.sendfile	= generic_file_sendfile,
1224 	.mmap		= ocfs2_mmap,
1225 	.fsync		= ocfs2_sync_file,
1226 	.release	= ocfs2_file_release,
1227 	.open		= ocfs2_file_open,
1228 	.aio_read	= ocfs2_file_aio_read,
1229 	.aio_write	= ocfs2_file_aio_write,
1230 };
1231 
1232 const struct file_operations ocfs2_dops = {
1233 	.read		= generic_read_dir,
1234 	.readdir	= ocfs2_readdir,
1235 	.fsync		= ocfs2_sync_file,
1236 };
1237