xref: /linux/fs/ocfs2/file.c (revision f24e9f586b377749dff37554696cf3a105540c94)
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 "ioctl.h"
48 #include "journal.h"
49 #include "mmap.h"
50 #include "suballoc.h"
51 #include "super.h"
52 
53 #include "buffer_head_io.h"
54 
55 static int ocfs2_sync_inode(struct inode *inode)
56 {
57 	filemap_fdatawrite(inode->i_mapping);
58 	return sync_mapping_buffers(inode->i_mapping);
59 }
60 
61 static int ocfs2_file_open(struct inode *inode, struct file *file)
62 {
63 	int status;
64 	int mode = file->f_flags;
65 	struct ocfs2_inode_info *oi = OCFS2_I(inode);
66 
67 	mlog_entry("(0x%p, 0x%p, '%.*s')\n", inode, file,
68 		   file->f_dentry->d_name.len, file->f_dentry->d_name.name);
69 
70 	spin_lock(&oi->ip_lock);
71 
72 	/* Check that the inode hasn't been wiped from disk by another
73 	 * node. If it hasn't then we're safe as long as we hold the
74 	 * spin lock until our increment of open count. */
75 	if (OCFS2_I(inode)->ip_flags & OCFS2_INODE_DELETED) {
76 		spin_unlock(&oi->ip_lock);
77 
78 		status = -ENOENT;
79 		goto leave;
80 	}
81 
82 	if (mode & O_DIRECT)
83 		oi->ip_flags |= OCFS2_INODE_OPEN_DIRECT;
84 
85 	oi->ip_open_count++;
86 	spin_unlock(&oi->ip_lock);
87 	status = 0;
88 leave:
89 	mlog_exit(status);
90 	return status;
91 }
92 
93 static int ocfs2_file_release(struct inode *inode, struct file *file)
94 {
95 	struct ocfs2_inode_info *oi = OCFS2_I(inode);
96 
97 	mlog_entry("(0x%p, 0x%p, '%.*s')\n", inode, file,
98 		       file->f_dentry->d_name.len,
99 		       file->f_dentry->d_name.name);
100 
101 	spin_lock(&oi->ip_lock);
102 	if (!--oi->ip_open_count)
103 		oi->ip_flags &= ~OCFS2_INODE_OPEN_DIRECT;
104 	spin_unlock(&oi->ip_lock);
105 
106 	mlog_exit(0);
107 
108 	return 0;
109 }
110 
111 static int ocfs2_sync_file(struct file *file,
112 			   struct dentry *dentry,
113 			   int datasync)
114 {
115 	int err = 0;
116 	journal_t *journal;
117 	struct inode *inode = dentry->d_inode;
118 	struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
119 
120 	mlog_entry("(0x%p, 0x%p, %d, '%.*s')\n", file, dentry, datasync,
121 		   dentry->d_name.len, dentry->d_name.name);
122 
123 	err = ocfs2_sync_inode(dentry->d_inode);
124 	if (err)
125 		goto bail;
126 
127 	journal = osb->journal->j_journal;
128 	err = journal_force_commit(journal);
129 
130 bail:
131 	mlog_exit(err);
132 
133 	return (err < 0) ? -EIO : 0;
134 }
135 
136 int ocfs2_set_inode_size(struct ocfs2_journal_handle *handle,
137 			 struct inode *inode,
138 			 struct buffer_head *fe_bh,
139 			 u64 new_i_size)
140 {
141 	int status;
142 
143 	mlog_entry_void();
144 	i_size_write(inode, new_i_size);
145 	inode->i_blocks = ocfs2_align_bytes_to_sectors(new_i_size);
146 	inode->i_ctime = inode->i_mtime = CURRENT_TIME;
147 
148 	status = ocfs2_mark_inode_dirty(handle, inode, fe_bh);
149 	if (status < 0) {
150 		mlog_errno(status);
151 		goto bail;
152 	}
153 
154 bail:
155 	mlog_exit(status);
156 	return status;
157 }
158 
159 static int ocfs2_simple_size_update(struct inode *inode,
160 				    struct buffer_head *di_bh,
161 				    u64 new_i_size)
162 {
163 	int ret;
164 	struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
165 	struct ocfs2_journal_handle *handle = NULL;
166 
167 	handle = ocfs2_start_trans(osb, NULL,
168 				   OCFS2_INODE_UPDATE_CREDITS);
169 	if (handle == NULL) {
170 		ret = -ENOMEM;
171 		mlog_errno(ret);
172 		goto out;
173 	}
174 
175 	ret = ocfs2_set_inode_size(handle, inode, di_bh,
176 				   new_i_size);
177 	if (ret < 0)
178 		mlog_errno(ret);
179 
180 	ocfs2_commit_trans(handle);
181 out:
182 	return ret;
183 }
184 
185 static int ocfs2_orphan_for_truncate(struct ocfs2_super *osb,
186 				     struct inode *inode,
187 				     struct buffer_head *fe_bh,
188 				     u64 new_i_size)
189 {
190 	int status;
191 	struct ocfs2_journal_handle *handle;
192 
193 	mlog_entry_void();
194 
195 	/* TODO: This needs to actually orphan the inode in this
196 	 * transaction. */
197 
198 	handle = ocfs2_start_trans(osb, NULL, OCFS2_INODE_UPDATE_CREDITS);
199 	if (IS_ERR(handle)) {
200 		status = PTR_ERR(handle);
201 		mlog_errno(status);
202 		goto out;
203 	}
204 
205 	status = ocfs2_set_inode_size(handle, inode, fe_bh, new_i_size);
206 	if (status < 0)
207 		mlog_errno(status);
208 
209 	ocfs2_commit_trans(handle);
210 out:
211 	mlog_exit(status);
212 	return status;
213 }
214 
215 static int ocfs2_truncate_file(struct inode *inode,
216 			       struct buffer_head *di_bh,
217 			       u64 new_i_size)
218 {
219 	int status = 0;
220 	struct ocfs2_dinode *fe = NULL;
221 	struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
222 	struct ocfs2_truncate_context *tc = NULL;
223 
224 	mlog_entry("(inode = %llu, new_i_size = %llu\n",
225 		   (unsigned long long)OCFS2_I(inode)->ip_blkno,
226 		   (unsigned long long)new_i_size);
227 
228 	truncate_inode_pages(inode->i_mapping, new_i_size);
229 
230 	fe = (struct ocfs2_dinode *) di_bh->b_data;
231 	if (!OCFS2_IS_VALID_DINODE(fe)) {
232 		OCFS2_RO_ON_INVALID_DINODE(inode->i_sb, fe);
233 		status = -EIO;
234 		goto bail;
235 	}
236 
237 	mlog_bug_on_msg(le64_to_cpu(fe->i_size) != i_size_read(inode),
238 			"Inode %llu, inode i_size = %lld != di "
239 			"i_size = %llu, i_flags = 0x%x\n",
240 			(unsigned long long)OCFS2_I(inode)->ip_blkno,
241 			i_size_read(inode),
242 			(unsigned long long)le64_to_cpu(fe->i_size),
243 			le32_to_cpu(fe->i_flags));
244 
245 	if (new_i_size > le64_to_cpu(fe->i_size)) {
246 		mlog(0, "asked to truncate file with size (%llu) to size (%llu)!\n",
247 		     (unsigned long long)le64_to_cpu(fe->i_size),
248 		     (unsigned long long)new_i_size);
249 		status = -EINVAL;
250 		mlog_errno(status);
251 		goto bail;
252 	}
253 
254 	mlog(0, "inode %llu, i_size = %llu, new_i_size = %llu\n",
255 	     (unsigned long long)le64_to_cpu(fe->i_blkno),
256 	     (unsigned long long)le64_to_cpu(fe->i_size),
257 	     (unsigned long long)new_i_size);
258 
259 	/* lets handle the simple truncate cases before doing any more
260 	 * cluster locking. */
261 	if (new_i_size == le64_to_cpu(fe->i_size))
262 		goto bail;
263 
264 	/* This forces other nodes to sync and drop their pages. Do
265 	 * this even if we have a truncate without allocation change -
266 	 * ocfs2 cluster sizes can be much greater than page size, so
267 	 * we have to truncate them anyway.  */
268 	status = ocfs2_data_lock(inode, 1);
269 	if (status < 0) {
270 		mlog_errno(status);
271 		goto bail;
272 	}
273 	ocfs2_data_unlock(inode, 1);
274 
275 	if (le32_to_cpu(fe->i_clusters) ==
276 	    ocfs2_clusters_for_bytes(osb->sb, new_i_size)) {
277 		mlog(0, "fe->i_clusters = %u, so we do a simple truncate\n",
278 		     fe->i_clusters);
279 		/* No allocation change is required, so lets fast path
280 		 * this truncate. */
281 		status = ocfs2_simple_size_update(inode, di_bh, new_i_size);
282 		if (status < 0)
283 			mlog_errno(status);
284 		goto bail;
285 	}
286 
287 	/* alright, we're going to need to do a full blown alloc size
288 	 * change. Orphan the inode so that recovery can complete the
289 	 * truncate if necessary. This does the task of marking
290 	 * i_size. */
291 	status = ocfs2_orphan_for_truncate(osb, inode, di_bh, new_i_size);
292 	if (status < 0) {
293 		mlog_errno(status);
294 		goto bail;
295 	}
296 
297 	status = ocfs2_prepare_truncate(osb, inode, di_bh, &tc);
298 	if (status < 0) {
299 		mlog_errno(status);
300 		goto bail;
301 	}
302 
303 	status = ocfs2_commit_truncate(osb, inode, di_bh, tc);
304 	if (status < 0) {
305 		mlog_errno(status);
306 		goto bail;
307 	}
308 
309 	/* TODO: orphan dir cleanup here. */
310 bail:
311 
312 	mlog_exit(status);
313 	return status;
314 }
315 
316 /*
317  * extend allocation only here.
318  * we'll update all the disk stuff, and oip->alloc_size
319  *
320  * expect stuff to be locked, a transaction started and enough data /
321  * metadata reservations in the contexts.
322  *
323  * Will return -EAGAIN, and a reason if a restart is needed.
324  * If passed in, *reason will always be set, even in error.
325  */
326 int ocfs2_do_extend_allocation(struct ocfs2_super *osb,
327 			       struct inode *inode,
328 			       u32 clusters_to_add,
329 			       struct buffer_head *fe_bh,
330 			       struct ocfs2_journal_handle *handle,
331 			       struct ocfs2_alloc_context *data_ac,
332 			       struct ocfs2_alloc_context *meta_ac,
333 			       enum ocfs2_alloc_restarted *reason_ret)
334 {
335 	int status = 0;
336 	int free_extents;
337 	struct ocfs2_dinode *fe = (struct ocfs2_dinode *) fe_bh->b_data;
338 	enum ocfs2_alloc_restarted reason = RESTART_NONE;
339 	u32 bit_off, num_bits;
340 	u64 block;
341 
342 	BUG_ON(!clusters_to_add);
343 
344 	free_extents = ocfs2_num_free_extents(osb, inode, fe);
345 	if (free_extents < 0) {
346 		status = free_extents;
347 		mlog_errno(status);
348 		goto leave;
349 	}
350 
351 	/* there are two cases which could cause us to EAGAIN in the
352 	 * we-need-more-metadata case:
353 	 * 1) we haven't reserved *any*
354 	 * 2) we are so fragmented, we've needed to add metadata too
355 	 *    many times. */
356 	if (!free_extents && !meta_ac) {
357 		mlog(0, "we haven't reserved any metadata!\n");
358 		status = -EAGAIN;
359 		reason = RESTART_META;
360 		goto leave;
361 	} else if ((!free_extents)
362 		   && (ocfs2_alloc_context_bits_left(meta_ac)
363 		       < ocfs2_extend_meta_needed(fe))) {
364 		mlog(0, "filesystem is really fragmented...\n");
365 		status = -EAGAIN;
366 		reason = RESTART_META;
367 		goto leave;
368 	}
369 
370 	status = ocfs2_claim_clusters(osb, handle, data_ac, 1,
371 				      &bit_off, &num_bits);
372 	if (status < 0) {
373 		if (status != -ENOSPC)
374 			mlog_errno(status);
375 		goto leave;
376 	}
377 
378 	BUG_ON(num_bits > clusters_to_add);
379 
380 	/* reserve our write early -- insert_extent may update the inode */
381 	status = ocfs2_journal_access(handle, inode, fe_bh,
382 				      OCFS2_JOURNAL_ACCESS_WRITE);
383 	if (status < 0) {
384 		mlog_errno(status);
385 		goto leave;
386 	}
387 
388 	block = ocfs2_clusters_to_blocks(osb->sb, bit_off);
389 	mlog(0, "Allocating %u clusters at block %u for inode %llu\n",
390 	     num_bits, bit_off, (unsigned long long)OCFS2_I(inode)->ip_blkno);
391 	status = ocfs2_insert_extent(osb, handle, inode, fe_bh, block,
392 				     num_bits, meta_ac);
393 	if (status < 0) {
394 		mlog_errno(status);
395 		goto leave;
396 	}
397 
398 	le32_add_cpu(&fe->i_clusters, num_bits);
399 	spin_lock(&OCFS2_I(inode)->ip_lock);
400 	OCFS2_I(inode)->ip_clusters = le32_to_cpu(fe->i_clusters);
401 	spin_unlock(&OCFS2_I(inode)->ip_lock);
402 
403 	status = ocfs2_journal_dirty(handle, fe_bh);
404 	if (status < 0) {
405 		mlog_errno(status);
406 		goto leave;
407 	}
408 
409 	clusters_to_add -= num_bits;
410 
411 	if (clusters_to_add) {
412 		mlog(0, "need to alloc once more, clusters = %u, wanted = "
413 		     "%u\n", fe->i_clusters, clusters_to_add);
414 		status = -EAGAIN;
415 		reason = RESTART_TRANS;
416 	}
417 
418 leave:
419 	mlog_exit(status);
420 	if (reason_ret)
421 		*reason_ret = reason;
422 	return status;
423 }
424 
425 static int ocfs2_extend_allocation(struct inode *inode,
426 				   u32 clusters_to_add)
427 {
428 	int status = 0;
429 	int restart_func = 0;
430 	int drop_alloc_sem = 0;
431 	int credits, num_free_extents;
432 	u32 prev_clusters;
433 	struct buffer_head *bh = NULL;
434 	struct ocfs2_dinode *fe = NULL;
435 	struct ocfs2_journal_handle *handle = NULL;
436 	struct ocfs2_alloc_context *data_ac = NULL;
437 	struct ocfs2_alloc_context *meta_ac = NULL;
438 	enum ocfs2_alloc_restarted why;
439 	struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
440 
441 	mlog_entry("(clusters_to_add = %u)\n", clusters_to_add);
442 
443 	status = ocfs2_read_block(osb, OCFS2_I(inode)->ip_blkno, &bh,
444 				  OCFS2_BH_CACHED, inode);
445 	if (status < 0) {
446 		mlog_errno(status);
447 		goto leave;
448 	}
449 
450 	fe = (struct ocfs2_dinode *) bh->b_data;
451 	if (!OCFS2_IS_VALID_DINODE(fe)) {
452 		OCFS2_RO_ON_INVALID_DINODE(inode->i_sb, fe);
453 		status = -EIO;
454 		goto leave;
455 	}
456 
457 restart_all:
458 	BUG_ON(le32_to_cpu(fe->i_clusters) != OCFS2_I(inode)->ip_clusters);
459 
460 	mlog(0, "extend inode %llu, i_size = %lld, fe->i_clusters = %u, "
461 	     "clusters_to_add = %u\n",
462 	     (unsigned long long)OCFS2_I(inode)->ip_blkno, i_size_read(inode),
463 	     fe->i_clusters, clusters_to_add);
464 
465 	handle = ocfs2_alloc_handle(osb);
466 	if (handle == NULL) {
467 		status = -ENOMEM;
468 		mlog_errno(status);
469 		goto leave;
470 	}
471 
472 	num_free_extents = ocfs2_num_free_extents(osb,
473 						  inode,
474 						  fe);
475 	if (num_free_extents < 0) {
476 		status = num_free_extents;
477 		mlog_errno(status);
478 		goto leave;
479 	}
480 
481 	if (!num_free_extents) {
482 		status = ocfs2_reserve_new_metadata(osb,
483 						    handle,
484 						    fe,
485 						    &meta_ac);
486 		if (status < 0) {
487 			if (status != -ENOSPC)
488 				mlog_errno(status);
489 			goto leave;
490 		}
491 	}
492 
493 	status = ocfs2_reserve_clusters(osb,
494 					handle,
495 					clusters_to_add,
496 					&data_ac);
497 	if (status < 0) {
498 		if (status != -ENOSPC)
499 			mlog_errno(status);
500 		goto leave;
501 	}
502 
503 	/* blocks peope in read/write from reading our allocation
504 	 * until we're done changing it. We depend on i_mutex to block
505 	 * other extend/truncate calls while we're here. Ordering wrt
506 	 * start_trans is important here -- always do it before! */
507 	down_write(&OCFS2_I(inode)->ip_alloc_sem);
508 	drop_alloc_sem = 1;
509 
510 	credits = ocfs2_calc_extend_credits(osb->sb, fe, clusters_to_add);
511 	handle = ocfs2_start_trans(osb, handle, credits);
512 	if (IS_ERR(handle)) {
513 		status = PTR_ERR(handle);
514 		handle = NULL;
515 		mlog_errno(status);
516 		goto leave;
517 	}
518 
519 restarted_transaction:
520 	/* reserve a write to the file entry early on - that we if we
521 	 * run out of credits in the allocation path, we can still
522 	 * update i_size. */
523 	status = ocfs2_journal_access(handle, inode, bh,
524 				      OCFS2_JOURNAL_ACCESS_WRITE);
525 	if (status < 0) {
526 		mlog_errno(status);
527 		goto leave;
528 	}
529 
530 	prev_clusters = OCFS2_I(inode)->ip_clusters;
531 
532 	status = ocfs2_do_extend_allocation(osb,
533 					    inode,
534 					    clusters_to_add,
535 					    bh,
536 					    handle,
537 					    data_ac,
538 					    meta_ac,
539 					    &why);
540 	if ((status < 0) && (status != -EAGAIN)) {
541 		if (status != -ENOSPC)
542 			mlog_errno(status);
543 		goto leave;
544 	}
545 
546 	status = ocfs2_journal_dirty(handle, bh);
547 	if (status < 0) {
548 		mlog_errno(status);
549 		goto leave;
550 	}
551 
552 	spin_lock(&OCFS2_I(inode)->ip_lock);
553 	clusters_to_add -= (OCFS2_I(inode)->ip_clusters - prev_clusters);
554 	spin_unlock(&OCFS2_I(inode)->ip_lock);
555 
556 	if (why != RESTART_NONE && clusters_to_add) {
557 		if (why == RESTART_META) {
558 			mlog(0, "restarting function.\n");
559 			restart_func = 1;
560 		} else {
561 			BUG_ON(why != RESTART_TRANS);
562 
563 			mlog(0, "restarting transaction.\n");
564 			/* TODO: This can be more intelligent. */
565 			credits = ocfs2_calc_extend_credits(osb->sb,
566 							    fe,
567 							    clusters_to_add);
568 			status = ocfs2_extend_trans(handle, credits);
569 			if (status < 0) {
570 				/* handle still has to be committed at
571 				 * this point. */
572 				status = -ENOMEM;
573 				mlog_errno(status);
574 				goto leave;
575 			}
576 			goto restarted_transaction;
577 		}
578 	}
579 
580 	mlog(0, "fe: i_clusters = %u, i_size=%llu\n",
581 	     fe->i_clusters, (unsigned long long)fe->i_size);
582 	mlog(0, "inode: ip_clusters=%u, i_size=%lld\n",
583 	     OCFS2_I(inode)->ip_clusters, i_size_read(inode));
584 
585 leave:
586 	if (drop_alloc_sem) {
587 		up_write(&OCFS2_I(inode)->ip_alloc_sem);
588 		drop_alloc_sem = 0;
589 	}
590 	if (handle) {
591 		ocfs2_commit_trans(handle);
592 		handle = NULL;
593 	}
594 	if (data_ac) {
595 		ocfs2_free_alloc_context(data_ac);
596 		data_ac = NULL;
597 	}
598 	if (meta_ac) {
599 		ocfs2_free_alloc_context(meta_ac);
600 		meta_ac = NULL;
601 	}
602 	if ((!status) && restart_func) {
603 		restart_func = 0;
604 		goto restart_all;
605 	}
606 	if (bh) {
607 		brelse(bh);
608 		bh = NULL;
609 	}
610 
611 	mlog_exit(status);
612 	return status;
613 }
614 
615 /* Some parts of this taken from generic_cont_expand, which turned out
616  * to be too fragile to do exactly what we need without us having to
617  * worry about recursive locking in ->prepare_write() and
618  * ->commit_write(). */
619 static int ocfs2_write_zero_page(struct inode *inode,
620 				 u64 size)
621 {
622 	struct address_space *mapping = inode->i_mapping;
623 	struct page *page;
624 	unsigned long index;
625 	unsigned int offset;
626 	struct ocfs2_journal_handle *handle = NULL;
627 	int ret;
628 
629 	offset = (size & (PAGE_CACHE_SIZE-1)); /* Within page */
630 	/* ugh.  in prepare/commit_write, if from==to==start of block, we
631 	** skip the prepare.  make sure we never send an offset for the start
632 	** of a block
633 	*/
634 	if ((offset & (inode->i_sb->s_blocksize - 1)) == 0) {
635 		offset++;
636 	}
637 	index = size >> PAGE_CACHE_SHIFT;
638 
639 	page = grab_cache_page(mapping, index);
640 	if (!page) {
641 		ret = -ENOMEM;
642 		mlog_errno(ret);
643 		goto out;
644 	}
645 
646 	ret = ocfs2_prepare_write_nolock(inode, page, offset, offset);
647 	if (ret < 0) {
648 		mlog_errno(ret);
649 		goto out_unlock;
650 	}
651 
652 	if (ocfs2_should_order_data(inode)) {
653 		handle = ocfs2_start_walk_page_trans(inode, page, offset,
654 						     offset);
655 		if (IS_ERR(handle)) {
656 			ret = PTR_ERR(handle);
657 			handle = NULL;
658 			goto out_unlock;
659 		}
660 	}
661 
662 	/* must not update i_size! */
663 	ret = block_commit_write(page, offset, offset);
664 	if (ret < 0)
665 		mlog_errno(ret);
666 	else
667 		ret = 0;
668 
669 	if (handle)
670 		ocfs2_commit_trans(handle);
671 out_unlock:
672 	unlock_page(page);
673 	page_cache_release(page);
674 out:
675 	return ret;
676 }
677 
678 static int ocfs2_zero_extend(struct inode *inode,
679 			     u64 zero_to_size)
680 {
681 	int ret = 0;
682 	u64 start_off;
683 	struct super_block *sb = inode->i_sb;
684 
685 	start_off = ocfs2_align_bytes_to_blocks(sb, i_size_read(inode));
686 	while (start_off < zero_to_size) {
687 		ret = ocfs2_write_zero_page(inode, start_off);
688 		if (ret < 0) {
689 			mlog_errno(ret);
690 			goto out;
691 		}
692 
693 		start_off += sb->s_blocksize;
694 	}
695 
696 out:
697 	return ret;
698 }
699 
700 /*
701  * A tail_to_skip value > 0 indicates that we're being called from
702  * ocfs2_file_aio_write(). This has the following implications:
703  *
704  * - we don't want to update i_size
705  * - di_bh will be NULL, which is fine because it's only used in the
706  *   case where we want to update i_size.
707  * - ocfs2_zero_extend() will then only be filling the hole created
708  *   between i_size and the start of the write.
709  */
710 static int ocfs2_extend_file(struct inode *inode,
711 			     struct buffer_head *di_bh,
712 			     u64 new_i_size,
713 			     size_t tail_to_skip)
714 {
715 	int ret = 0;
716 	u32 clusters_to_add;
717 
718 	BUG_ON(!tail_to_skip && !di_bh);
719 
720 	/* setattr sometimes calls us like this. */
721 	if (new_i_size == 0)
722 		goto out;
723 
724 	if (i_size_read(inode) == new_i_size)
725   		goto out;
726 	BUG_ON(new_i_size < i_size_read(inode));
727 
728 	clusters_to_add = ocfs2_clusters_for_bytes(inode->i_sb, new_i_size) -
729 		OCFS2_I(inode)->ip_clusters;
730 
731 	if (clusters_to_add) {
732 		/*
733 		 * protect the pages that ocfs2_zero_extend is going to
734 		 * be pulling into the page cache.. we do this before the
735 		 * metadata extend so that we don't get into the situation
736 		 * where we've extended the metadata but can't get the data
737 		 * lock to zero.
738 		 */
739 		ret = ocfs2_data_lock(inode, 1);
740 		if (ret < 0) {
741 			mlog_errno(ret);
742 			goto out;
743 		}
744 
745 		ret = ocfs2_extend_allocation(inode, clusters_to_add);
746 		if (ret < 0) {
747 			mlog_errno(ret);
748 			goto out_unlock;
749 		}
750 
751 		ret = ocfs2_zero_extend(inode, (u64)new_i_size - tail_to_skip);
752 		if (ret < 0) {
753 			mlog_errno(ret);
754 			goto out_unlock;
755 		}
756 	}
757 
758 	if (!tail_to_skip) {
759 		/* We're being called from ocfs2_setattr() which wants
760 		 * us to update i_size */
761 		ret = ocfs2_simple_size_update(inode, di_bh, new_i_size);
762 		if (ret < 0)
763 			mlog_errno(ret);
764 	}
765 
766 out_unlock:
767 	if (clusters_to_add) /* this is the only case in which we lock */
768 		ocfs2_data_unlock(inode, 1);
769 
770 out:
771 	return ret;
772 }
773 
774 int ocfs2_setattr(struct dentry *dentry, struct iattr *attr)
775 {
776 	int status = 0, size_change;
777 	struct inode *inode = dentry->d_inode;
778 	struct super_block *sb = inode->i_sb;
779 	struct ocfs2_super *osb = OCFS2_SB(sb);
780 	struct buffer_head *bh = NULL;
781 	struct ocfs2_journal_handle *handle = NULL;
782 
783 	mlog_entry("(0x%p, '%.*s')\n", dentry,
784 	           dentry->d_name.len, dentry->d_name.name);
785 
786 	if (attr->ia_valid & ATTR_MODE)
787 		mlog(0, "mode change: %d\n", attr->ia_mode);
788 	if (attr->ia_valid & ATTR_UID)
789 		mlog(0, "uid change: %d\n", attr->ia_uid);
790 	if (attr->ia_valid & ATTR_GID)
791 		mlog(0, "gid change: %d\n", attr->ia_gid);
792 	if (attr->ia_valid & ATTR_SIZE)
793 		mlog(0, "size change...\n");
794 	if (attr->ia_valid & (ATTR_ATIME | ATTR_MTIME | ATTR_CTIME))
795 		mlog(0, "time change...\n");
796 
797 #define OCFS2_VALID_ATTRS (ATTR_ATIME | ATTR_MTIME | ATTR_CTIME | ATTR_SIZE \
798 			   | ATTR_GID | ATTR_UID | ATTR_MODE)
799 	if (!(attr->ia_valid & OCFS2_VALID_ATTRS)) {
800 		mlog(0, "can't handle attrs: 0x%x\n", attr->ia_valid);
801 		return 0;
802 	}
803 
804 	status = inode_change_ok(inode, attr);
805 	if (status)
806 		return status;
807 
808 	size_change = S_ISREG(inode->i_mode) && attr->ia_valid & ATTR_SIZE;
809 	if (size_change) {
810 		status = ocfs2_rw_lock(inode, 1);
811 		if (status < 0) {
812 			mlog_errno(status);
813 			goto bail;
814 		}
815 	}
816 
817 	status = ocfs2_meta_lock(inode, NULL, &bh, 1);
818 	if (status < 0) {
819 		if (status != -ENOENT)
820 			mlog_errno(status);
821 		goto bail_unlock_rw;
822 	}
823 
824 	if (size_change && attr->ia_size != i_size_read(inode)) {
825 		if (i_size_read(inode) > attr->ia_size)
826 			status = ocfs2_truncate_file(inode, bh, attr->ia_size);
827 		else
828 			status = ocfs2_extend_file(inode, bh, attr->ia_size, 0);
829 		if (status < 0) {
830 			if (status != -ENOSPC)
831 				mlog_errno(status);
832 			status = -ENOSPC;
833 			goto bail_unlock;
834 		}
835 	}
836 
837 	handle = ocfs2_start_trans(osb, NULL, OCFS2_INODE_UPDATE_CREDITS);
838 	if (IS_ERR(handle)) {
839 		status = PTR_ERR(handle);
840 		mlog_errno(status);
841 		goto bail_unlock;
842 	}
843 
844 	status = inode_setattr(inode, attr);
845 	if (status < 0) {
846 		mlog_errno(status);
847 		goto bail_commit;
848 	}
849 
850 	status = ocfs2_mark_inode_dirty(handle, inode, bh);
851 	if (status < 0)
852 		mlog_errno(status);
853 
854 bail_commit:
855 	ocfs2_commit_trans(handle);
856 bail_unlock:
857 	ocfs2_meta_unlock(inode, 1);
858 bail_unlock_rw:
859 	if (size_change)
860 		ocfs2_rw_unlock(inode, 1);
861 bail:
862 	if (bh)
863 		brelse(bh);
864 
865 	mlog_exit(status);
866 	return status;
867 }
868 
869 int ocfs2_getattr(struct vfsmount *mnt,
870 		  struct dentry *dentry,
871 		  struct kstat *stat)
872 {
873 	struct inode *inode = dentry->d_inode;
874 	struct super_block *sb = dentry->d_inode->i_sb;
875 	struct ocfs2_super *osb = sb->s_fs_info;
876 	int err;
877 
878 	mlog_entry_void();
879 
880 	err = ocfs2_inode_revalidate(dentry);
881 	if (err) {
882 		if (err != -ENOENT)
883 			mlog_errno(err);
884 		goto bail;
885 	}
886 
887 	generic_fillattr(inode, stat);
888 
889 	/* We set the blksize from the cluster size for performance */
890 	stat->blksize = osb->s_clustersize;
891 
892 bail:
893 	mlog_exit(err);
894 
895 	return err;
896 }
897 
898 static int ocfs2_write_remove_suid(struct inode *inode)
899 {
900 	int ret;
901 	struct buffer_head *bh = NULL;
902 	struct ocfs2_inode_info *oi = OCFS2_I(inode);
903 	struct ocfs2_journal_handle *handle;
904 	struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
905 	struct ocfs2_dinode *di;
906 
907 	mlog_entry("(Inode %llu, mode 0%o)\n",
908 		   (unsigned long long)oi->ip_blkno, inode->i_mode);
909 
910 	handle = ocfs2_start_trans(osb, NULL, OCFS2_INODE_UPDATE_CREDITS);
911 	if (handle == NULL) {
912 		ret = -ENOMEM;
913 		mlog_errno(ret);
914 		goto out;
915 	}
916 
917 	ret = ocfs2_read_block(osb, oi->ip_blkno, &bh, OCFS2_BH_CACHED, inode);
918 	if (ret < 0) {
919 		mlog_errno(ret);
920 		goto out_trans;
921 	}
922 
923 	ret = ocfs2_journal_access(handle, inode, bh,
924 				   OCFS2_JOURNAL_ACCESS_WRITE);
925 	if (ret < 0) {
926 		mlog_errno(ret);
927 		goto out_bh;
928 	}
929 
930 	inode->i_mode &= ~S_ISUID;
931 	if ((inode->i_mode & S_ISGID) && (inode->i_mode & S_IXGRP))
932 		inode->i_mode &= ~S_ISGID;
933 
934 	di = (struct ocfs2_dinode *) bh->b_data;
935 	di->i_mode = cpu_to_le16(inode->i_mode);
936 
937 	ret = ocfs2_journal_dirty(handle, bh);
938 	if (ret < 0)
939 		mlog_errno(ret);
940 out_bh:
941 	brelse(bh);
942 out_trans:
943 	ocfs2_commit_trans(handle);
944 out:
945 	mlog_exit(ret);
946 	return ret;
947 }
948 
949 static inline int ocfs2_write_should_remove_suid(struct inode *inode)
950 {
951 	mode_t mode = inode->i_mode;
952 
953 	if (!capable(CAP_FSETID)) {
954 		if (unlikely(mode & S_ISUID))
955 			return 1;
956 
957 		if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
958 			return 1;
959 	}
960 	return 0;
961 }
962 
963 static ssize_t ocfs2_file_aio_write(struct kiocb *iocb,
964 				    const char __user *buf,
965 				    size_t count,
966 				    loff_t pos)
967 {
968 	struct iovec local_iov = { .iov_base = (void __user *)buf,
969 				   .iov_len = count };
970 	int ret, rw_level = -1, meta_level = -1, have_alloc_sem = 0;
971 	u32 clusters;
972 	struct file *filp = iocb->ki_filp;
973 	struct inode *inode = filp->f_dentry->d_inode;
974 	loff_t newsize, saved_pos;
975 
976 	mlog_entry("(0x%p, 0x%p, %u, '%.*s')\n", filp, buf,
977 		   (unsigned int)count,
978 		   filp->f_dentry->d_name.len,
979 		   filp->f_dentry->d_name.name);
980 
981 	/* happy write of zero bytes */
982 	if (count == 0)
983 		return 0;
984 
985 	if (!inode) {
986 		mlog(0, "bad inode\n");
987 		return -EIO;
988 	}
989 
990 	mutex_lock(&inode->i_mutex);
991 	/* to match setattr's i_mutex -> i_alloc_sem -> rw_lock ordering */
992 	if (filp->f_flags & O_DIRECT) {
993 		have_alloc_sem = 1;
994 		down_read(&inode->i_alloc_sem);
995 	}
996 
997 	/* concurrent O_DIRECT writes are allowed */
998 	rw_level = (filp->f_flags & O_DIRECT) ? 0 : 1;
999 	ret = ocfs2_rw_lock(inode, rw_level);
1000 	if (ret < 0) {
1001 		rw_level = -1;
1002 		mlog_errno(ret);
1003 		goto out;
1004 	}
1005 
1006 	/*
1007 	 * We sample i_size under a read level meta lock to see if our write
1008 	 * is extending the file, if it is we back off and get a write level
1009 	 * meta lock.
1010 	 */
1011 	meta_level = (filp->f_flags & O_APPEND) ? 1 : 0;
1012 	for(;;) {
1013 		ret = ocfs2_meta_lock(inode, NULL, NULL, meta_level);
1014 		if (ret < 0) {
1015 			meta_level = -1;
1016 			mlog_errno(ret);
1017 			goto out;
1018 		}
1019 
1020 		/* Clear suid / sgid if necessary. We do this here
1021 		 * instead of later in the write path because
1022 		 * remove_suid() calls ->setattr without any hint that
1023 		 * we may have already done our cluster locking. Since
1024 		 * ocfs2_setattr() *must* take cluster locks to
1025 		 * proceeed, this will lead us to recursively lock the
1026 		 * inode. There's also the dinode i_size state which
1027 		 * can be lost via setattr during extending writes (we
1028 		 * set inode->i_size at the end of a write. */
1029 		if (ocfs2_write_should_remove_suid(inode)) {
1030 			if (meta_level == 0) {
1031 				ocfs2_meta_unlock(inode, meta_level);
1032 				meta_level = 1;
1033 				continue;
1034 			}
1035 
1036 			ret = ocfs2_write_remove_suid(inode);
1037 			if (ret < 0) {
1038 				mlog_errno(ret);
1039 				goto out;
1040 			}
1041 		}
1042 
1043 		/* work on a copy of ppos until we're sure that we won't have
1044 		 * to recalculate it due to relocking. */
1045 		if (filp->f_flags & O_APPEND) {
1046 			saved_pos = i_size_read(inode);
1047 			mlog(0, "O_APPEND: inode->i_size=%llu\n", saved_pos);
1048 		} else {
1049 			saved_pos = iocb->ki_pos;
1050 		}
1051 		newsize = count + saved_pos;
1052 
1053 		mlog(0, "pos=%lld newsize=%lld cursize=%lld\n",
1054 		     (long long) saved_pos, (long long) newsize,
1055 		     (long long) i_size_read(inode));
1056 
1057 		/* No need for a higher level metadata lock if we're
1058 		 * never going past i_size. */
1059 		if (newsize <= i_size_read(inode))
1060 			break;
1061 
1062 		if (meta_level == 0) {
1063 			ocfs2_meta_unlock(inode, meta_level);
1064 			meta_level = 1;
1065 			continue;
1066 		}
1067 
1068 		spin_lock(&OCFS2_I(inode)->ip_lock);
1069 		clusters = ocfs2_clusters_for_bytes(inode->i_sb, newsize) -
1070 			OCFS2_I(inode)->ip_clusters;
1071 		spin_unlock(&OCFS2_I(inode)->ip_lock);
1072 
1073 		mlog(0, "Writing at EOF, may need more allocation: "
1074 		     "i_size = %lld, newsize = %lld, need %u clusters\n",
1075 		     (long long) i_size_read(inode), (long long) newsize,
1076 		     clusters);
1077 
1078 		/* We only want to continue the rest of this loop if
1079 		 * our extend will actually require more
1080 		 * allocation. */
1081 		if (!clusters)
1082 			break;
1083 
1084 		ret = ocfs2_extend_file(inode, NULL, newsize, count);
1085 		if (ret < 0) {
1086 			if (ret != -ENOSPC)
1087 				mlog_errno(ret);
1088 			goto out;
1089 		}
1090 		break;
1091 	}
1092 
1093 	/* ok, we're done with i_size and alloc work */
1094 	iocb->ki_pos = saved_pos;
1095 	ocfs2_meta_unlock(inode, meta_level);
1096 	meta_level = -1;
1097 
1098 	/* communicate with ocfs2_dio_end_io */
1099 	ocfs2_iocb_set_rw_locked(iocb);
1100 
1101 	ret = generic_file_aio_write_nolock(iocb, &local_iov, 1, &iocb->ki_pos);
1102 
1103 	/* buffered aio wouldn't have proper lock coverage today */
1104 	BUG_ON(ret == -EIOCBQUEUED && !(filp->f_flags & O_DIRECT));
1105 
1106 	/*
1107 	 * deep in g_f_a_w_n()->ocfs2_direct_IO we pass in a ocfs2_dio_end_io
1108 	 * function pointer which is called when o_direct io completes so that
1109 	 * it can unlock our rw lock.  (it's the clustered equivalent of
1110 	 * i_alloc_sem; protects truncate from racing with pending ios).
1111 	 * Unfortunately there are error cases which call end_io and others
1112 	 * that don't.  so we don't have to unlock the rw_lock if either an
1113 	 * async dio is going to do it in the future or an end_io after an
1114 	 * error has already done it.
1115 	 */
1116 	if (ret == -EIOCBQUEUED || !ocfs2_iocb_is_rw_locked(iocb)) {
1117 		rw_level = -1;
1118 		have_alloc_sem = 0;
1119 	}
1120 
1121 out:
1122 	if (meta_level != -1)
1123 		ocfs2_meta_unlock(inode, meta_level);
1124 	if (have_alloc_sem)
1125 		up_read(&inode->i_alloc_sem);
1126 	if (rw_level != -1)
1127 		ocfs2_rw_unlock(inode, rw_level);
1128 	mutex_unlock(&inode->i_mutex);
1129 
1130 	mlog_exit(ret);
1131 	return ret;
1132 }
1133 
1134 static ssize_t ocfs2_file_aio_read(struct kiocb *iocb,
1135 				   char __user *buf,
1136 				   size_t count,
1137 				   loff_t pos)
1138 {
1139 	int ret = 0, rw_level = -1, have_alloc_sem = 0;
1140 	struct file *filp = iocb->ki_filp;
1141 	struct inode *inode = filp->f_dentry->d_inode;
1142 
1143 	mlog_entry("(0x%p, 0x%p, %u, '%.*s')\n", filp, buf,
1144 		   (unsigned int)count,
1145 		   filp->f_dentry->d_name.len,
1146 		   filp->f_dentry->d_name.name);
1147 
1148 	if (!inode) {
1149 		ret = -EINVAL;
1150 		mlog_errno(ret);
1151 		goto bail;
1152 	}
1153 
1154 	/*
1155 	 * buffered reads protect themselves in ->readpage().  O_DIRECT reads
1156 	 * need locks to protect pending reads from racing with truncate.
1157 	 */
1158 	if (filp->f_flags & O_DIRECT) {
1159 		down_read(&inode->i_alloc_sem);
1160 		have_alloc_sem = 1;
1161 
1162 		ret = ocfs2_rw_lock(inode, 0);
1163 		if (ret < 0) {
1164 			mlog_errno(ret);
1165 			goto bail;
1166 		}
1167 		rw_level = 0;
1168 		/* communicate with ocfs2_dio_end_io */
1169 		ocfs2_iocb_set_rw_locked(iocb);
1170 	}
1171 
1172 	/*
1173 	 * We're fine letting folks race truncates and extending
1174 	 * writes with read across the cluster, just like they can
1175 	 * locally. Hence no rw_lock during read.
1176 	 *
1177 	 * Take and drop the meta data lock to update inode fields
1178 	 * like i_size. This allows the checks down below
1179 	 * generic_file_aio_read() a chance of actually working.
1180 	 */
1181 	ret = ocfs2_meta_lock(inode, NULL, NULL, 0);
1182 	if (ret < 0) {
1183 		mlog_errno(ret);
1184 		goto bail;
1185 	}
1186 	ocfs2_meta_unlock(inode, 0);
1187 
1188 	ret = generic_file_aio_read(iocb, buf, count, iocb->ki_pos);
1189 	if (ret == -EINVAL)
1190 		mlog(ML_ERROR, "generic_file_aio_read returned -EINVAL\n");
1191 
1192 	/* buffered aio wouldn't have proper lock coverage today */
1193 	BUG_ON(ret == -EIOCBQUEUED && !(filp->f_flags & O_DIRECT));
1194 
1195 	/* see ocfs2_file_aio_write */
1196 	if (ret == -EIOCBQUEUED || !ocfs2_iocb_is_rw_locked(iocb)) {
1197 		rw_level = -1;
1198 		have_alloc_sem = 0;
1199 	}
1200 
1201 bail:
1202 	if (have_alloc_sem)
1203 		up_read(&inode->i_alloc_sem);
1204 	if (rw_level != -1)
1205 		ocfs2_rw_unlock(inode, rw_level);
1206 	mlog_exit(ret);
1207 
1208 	return ret;
1209 }
1210 
1211 struct inode_operations ocfs2_file_iops = {
1212 	.setattr	= ocfs2_setattr,
1213 	.getattr	= ocfs2_getattr,
1214 };
1215 
1216 struct inode_operations ocfs2_special_file_iops = {
1217 	.setattr	= ocfs2_setattr,
1218 	.getattr	= ocfs2_getattr,
1219 };
1220 
1221 const struct file_operations ocfs2_fops = {
1222 	.read		= do_sync_read,
1223 	.write		= do_sync_write,
1224 	.sendfile	= generic_file_sendfile,
1225 	.mmap		= ocfs2_mmap,
1226 	.fsync		= ocfs2_sync_file,
1227 	.release	= ocfs2_file_release,
1228 	.open		= ocfs2_file_open,
1229 	.aio_read	= ocfs2_file_aio_read,
1230 	.aio_write	= ocfs2_file_aio_write,
1231 	.ioctl		= ocfs2_ioctl,
1232 };
1233 
1234 const struct file_operations ocfs2_dops = {
1235 	.read		= generic_read_dir,
1236 	.readdir	= ocfs2_readdir,
1237 	.fsync		= ocfs2_sync_file,
1238 	.ioctl		= ocfs2_ioctl,
1239 };
1240