xref: /openbmc/linux/fs/ocfs2/aops.c (revision 74ba9207)
1 /* -*- mode: c; c-basic-offset: 8; -*-
2  * vim: noexpandtab sw=8 ts=8 sts=0:
3  *
4  * Copyright (C) 2002, 2004 Oracle.  All rights reserved.
5  *
6  * This program is free software; you can redistribute it and/or
7  * modify it under the terms of the GNU General Public
8  * License as published by the Free Software Foundation; either
9  * version 2 of the License, or (at your option) any later version.
10  *
11  * This program is distributed in the hope that it will be useful,
12  * but WITHOUT ANY WARRANTY; without even the implied warranty of
13  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
14  * General Public License for more details.
15  *
16  * You should have received a copy of the GNU General Public
17  * License along with this program; if not, write to the
18  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
19  * Boston, MA 021110-1307, USA.
20  */
21 
22 #include <linux/fs.h>
23 #include <linux/slab.h>
24 #include <linux/highmem.h>
25 #include <linux/pagemap.h>
26 #include <asm/byteorder.h>
27 #include <linux/swap.h>
28 #include <linux/pipe_fs_i.h>
29 #include <linux/mpage.h>
30 #include <linux/quotaops.h>
31 #include <linux/blkdev.h>
32 #include <linux/uio.h>
33 #include <linux/mm.h>
34 
35 #include <cluster/masklog.h>
36 
37 #include "ocfs2.h"
38 
39 #include "alloc.h"
40 #include "aops.h"
41 #include "dlmglue.h"
42 #include "extent_map.h"
43 #include "file.h"
44 #include "inode.h"
45 #include "journal.h"
46 #include "suballoc.h"
47 #include "super.h"
48 #include "symlink.h"
49 #include "refcounttree.h"
50 #include "ocfs2_trace.h"
51 
52 #include "buffer_head_io.h"
53 #include "dir.h"
54 #include "namei.h"
55 #include "sysfile.h"
56 
57 static int ocfs2_symlink_get_block(struct inode *inode, sector_t iblock,
58 				   struct buffer_head *bh_result, int create)
59 {
60 	int err = -EIO;
61 	int status;
62 	struct ocfs2_dinode *fe = NULL;
63 	struct buffer_head *bh = NULL;
64 	struct buffer_head *buffer_cache_bh = NULL;
65 	struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
66 	void *kaddr;
67 
68 	trace_ocfs2_symlink_get_block(
69 			(unsigned long long)OCFS2_I(inode)->ip_blkno,
70 			(unsigned long long)iblock, bh_result, create);
71 
72 	BUG_ON(ocfs2_inode_is_fast_symlink(inode));
73 
74 	if ((iblock << inode->i_sb->s_blocksize_bits) > PATH_MAX + 1) {
75 		mlog(ML_ERROR, "block offset > PATH_MAX: %llu",
76 		     (unsigned long long)iblock);
77 		goto bail;
78 	}
79 
80 	status = ocfs2_read_inode_block(inode, &bh);
81 	if (status < 0) {
82 		mlog_errno(status);
83 		goto bail;
84 	}
85 	fe = (struct ocfs2_dinode *) bh->b_data;
86 
87 	if ((u64)iblock >= ocfs2_clusters_to_blocks(inode->i_sb,
88 						    le32_to_cpu(fe->i_clusters))) {
89 		err = -ENOMEM;
90 		mlog(ML_ERROR, "block offset is outside the allocated size: "
91 		     "%llu\n", (unsigned long long)iblock);
92 		goto bail;
93 	}
94 
95 	/* We don't use the page cache to create symlink data, so if
96 	 * need be, copy it over from the buffer cache. */
97 	if (!buffer_uptodate(bh_result) && ocfs2_inode_is_new(inode)) {
98 		u64 blkno = le64_to_cpu(fe->id2.i_list.l_recs[0].e_blkno) +
99 			    iblock;
100 		buffer_cache_bh = sb_getblk(osb->sb, blkno);
101 		if (!buffer_cache_bh) {
102 			err = -ENOMEM;
103 			mlog(ML_ERROR, "couldn't getblock for symlink!\n");
104 			goto bail;
105 		}
106 
107 		/* we haven't locked out transactions, so a commit
108 		 * could've happened. Since we've got a reference on
109 		 * the bh, even if it commits while we're doing the
110 		 * copy, the data is still good. */
111 		if (buffer_jbd(buffer_cache_bh)
112 		    && ocfs2_inode_is_new(inode)) {
113 			kaddr = kmap_atomic(bh_result->b_page);
114 			if (!kaddr) {
115 				mlog(ML_ERROR, "couldn't kmap!\n");
116 				goto bail;
117 			}
118 			memcpy(kaddr + (bh_result->b_size * iblock),
119 			       buffer_cache_bh->b_data,
120 			       bh_result->b_size);
121 			kunmap_atomic(kaddr);
122 			set_buffer_uptodate(bh_result);
123 		}
124 		brelse(buffer_cache_bh);
125 	}
126 
127 	map_bh(bh_result, inode->i_sb,
128 	       le64_to_cpu(fe->id2.i_list.l_recs[0].e_blkno) + iblock);
129 
130 	err = 0;
131 
132 bail:
133 	brelse(bh);
134 
135 	return err;
136 }
137 
138 static int ocfs2_lock_get_block(struct inode *inode, sector_t iblock,
139 		    struct buffer_head *bh_result, int create)
140 {
141 	int ret = 0;
142 	struct ocfs2_inode_info *oi = OCFS2_I(inode);
143 
144 	down_read(&oi->ip_alloc_sem);
145 	ret = ocfs2_get_block(inode, iblock, bh_result, create);
146 	up_read(&oi->ip_alloc_sem);
147 
148 	return ret;
149 }
150 
151 int ocfs2_get_block(struct inode *inode, sector_t iblock,
152 		    struct buffer_head *bh_result, int create)
153 {
154 	int err = 0;
155 	unsigned int ext_flags;
156 	u64 max_blocks = bh_result->b_size >> inode->i_blkbits;
157 	u64 p_blkno, count, past_eof;
158 	struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
159 
160 	trace_ocfs2_get_block((unsigned long long)OCFS2_I(inode)->ip_blkno,
161 			      (unsigned long long)iblock, bh_result, create);
162 
163 	if (OCFS2_I(inode)->ip_flags & OCFS2_INODE_SYSTEM_FILE)
164 		mlog(ML_NOTICE, "get_block on system inode 0x%p (%lu)\n",
165 		     inode, inode->i_ino);
166 
167 	if (S_ISLNK(inode->i_mode)) {
168 		/* this always does I/O for some reason. */
169 		err = ocfs2_symlink_get_block(inode, iblock, bh_result, create);
170 		goto bail;
171 	}
172 
173 	err = ocfs2_extent_map_get_blocks(inode, iblock, &p_blkno, &count,
174 					  &ext_flags);
175 	if (err) {
176 		mlog(ML_ERROR, "Error %d from get_blocks(0x%p, %llu, 1, "
177 		     "%llu, NULL)\n", err, inode, (unsigned long long)iblock,
178 		     (unsigned long long)p_blkno);
179 		goto bail;
180 	}
181 
182 	if (max_blocks < count)
183 		count = max_blocks;
184 
185 	/*
186 	 * ocfs2 never allocates in this function - the only time we
187 	 * need to use BH_New is when we're extending i_size on a file
188 	 * system which doesn't support holes, in which case BH_New
189 	 * allows __block_write_begin() to zero.
190 	 *
191 	 * If we see this on a sparse file system, then a truncate has
192 	 * raced us and removed the cluster. In this case, we clear
193 	 * the buffers dirty and uptodate bits and let the buffer code
194 	 * ignore it as a hole.
195 	 */
196 	if (create && p_blkno == 0 && ocfs2_sparse_alloc(osb)) {
197 		clear_buffer_dirty(bh_result);
198 		clear_buffer_uptodate(bh_result);
199 		goto bail;
200 	}
201 
202 	/* Treat the unwritten extent as a hole for zeroing purposes. */
203 	if (p_blkno && !(ext_flags & OCFS2_EXT_UNWRITTEN))
204 		map_bh(bh_result, inode->i_sb, p_blkno);
205 
206 	bh_result->b_size = count << inode->i_blkbits;
207 
208 	if (!ocfs2_sparse_alloc(osb)) {
209 		if (p_blkno == 0) {
210 			err = -EIO;
211 			mlog(ML_ERROR,
212 			     "iblock = %llu p_blkno = %llu blkno=(%llu)\n",
213 			     (unsigned long long)iblock,
214 			     (unsigned long long)p_blkno,
215 			     (unsigned long long)OCFS2_I(inode)->ip_blkno);
216 			mlog(ML_ERROR, "Size %llu, clusters %u\n", (unsigned long long)i_size_read(inode), OCFS2_I(inode)->ip_clusters);
217 			dump_stack();
218 			goto bail;
219 		}
220 	}
221 
222 	past_eof = ocfs2_blocks_for_bytes(inode->i_sb, i_size_read(inode));
223 
224 	trace_ocfs2_get_block_end((unsigned long long)OCFS2_I(inode)->ip_blkno,
225 				  (unsigned long long)past_eof);
226 	if (create && (iblock >= past_eof))
227 		set_buffer_new(bh_result);
228 
229 bail:
230 	if (err < 0)
231 		err = -EIO;
232 
233 	return err;
234 }
235 
236 int ocfs2_read_inline_data(struct inode *inode, struct page *page,
237 			   struct buffer_head *di_bh)
238 {
239 	void *kaddr;
240 	loff_t size;
241 	struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
242 
243 	if (!(le16_to_cpu(di->i_dyn_features) & OCFS2_INLINE_DATA_FL)) {
244 		ocfs2_error(inode->i_sb, "Inode %llu lost inline data flag\n",
245 			    (unsigned long long)OCFS2_I(inode)->ip_blkno);
246 		return -EROFS;
247 	}
248 
249 	size = i_size_read(inode);
250 
251 	if (size > PAGE_SIZE ||
252 	    size > ocfs2_max_inline_data_with_xattr(inode->i_sb, di)) {
253 		ocfs2_error(inode->i_sb,
254 			    "Inode %llu has with inline data has bad size: %Lu\n",
255 			    (unsigned long long)OCFS2_I(inode)->ip_blkno,
256 			    (unsigned long long)size);
257 		return -EROFS;
258 	}
259 
260 	kaddr = kmap_atomic(page);
261 	if (size)
262 		memcpy(kaddr, di->id2.i_data.id_data, size);
263 	/* Clear the remaining part of the page */
264 	memset(kaddr + size, 0, PAGE_SIZE - size);
265 	flush_dcache_page(page);
266 	kunmap_atomic(kaddr);
267 
268 	SetPageUptodate(page);
269 
270 	return 0;
271 }
272 
273 static int ocfs2_readpage_inline(struct inode *inode, struct page *page)
274 {
275 	int ret;
276 	struct buffer_head *di_bh = NULL;
277 
278 	BUG_ON(!PageLocked(page));
279 	BUG_ON(!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL));
280 
281 	ret = ocfs2_read_inode_block(inode, &di_bh);
282 	if (ret) {
283 		mlog_errno(ret);
284 		goto out;
285 	}
286 
287 	ret = ocfs2_read_inline_data(inode, page, di_bh);
288 out:
289 	unlock_page(page);
290 
291 	brelse(di_bh);
292 	return ret;
293 }
294 
295 static int ocfs2_readpage(struct file *file, struct page *page)
296 {
297 	struct inode *inode = page->mapping->host;
298 	struct ocfs2_inode_info *oi = OCFS2_I(inode);
299 	loff_t start = (loff_t)page->index << PAGE_SHIFT;
300 	int ret, unlock = 1;
301 
302 	trace_ocfs2_readpage((unsigned long long)oi->ip_blkno,
303 			     (page ? page->index : 0));
304 
305 	ret = ocfs2_inode_lock_with_page(inode, NULL, 0, page);
306 	if (ret != 0) {
307 		if (ret == AOP_TRUNCATED_PAGE)
308 			unlock = 0;
309 		mlog_errno(ret);
310 		goto out;
311 	}
312 
313 	if (down_read_trylock(&oi->ip_alloc_sem) == 0) {
314 		/*
315 		 * Unlock the page and cycle ip_alloc_sem so that we don't
316 		 * busyloop waiting for ip_alloc_sem to unlock
317 		 */
318 		ret = AOP_TRUNCATED_PAGE;
319 		unlock_page(page);
320 		unlock = 0;
321 		down_read(&oi->ip_alloc_sem);
322 		up_read(&oi->ip_alloc_sem);
323 		goto out_inode_unlock;
324 	}
325 
326 	/*
327 	 * i_size might have just been updated as we grabed the meta lock.  We
328 	 * might now be discovering a truncate that hit on another node.
329 	 * block_read_full_page->get_block freaks out if it is asked to read
330 	 * beyond the end of a file, so we check here.  Callers
331 	 * (generic_file_read, vm_ops->fault) are clever enough to check i_size
332 	 * and notice that the page they just read isn't needed.
333 	 *
334 	 * XXX sys_readahead() seems to get that wrong?
335 	 */
336 	if (start >= i_size_read(inode)) {
337 		zero_user(page, 0, PAGE_SIZE);
338 		SetPageUptodate(page);
339 		ret = 0;
340 		goto out_alloc;
341 	}
342 
343 	if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL)
344 		ret = ocfs2_readpage_inline(inode, page);
345 	else
346 		ret = block_read_full_page(page, ocfs2_get_block);
347 	unlock = 0;
348 
349 out_alloc:
350 	up_read(&oi->ip_alloc_sem);
351 out_inode_unlock:
352 	ocfs2_inode_unlock(inode, 0);
353 out:
354 	if (unlock)
355 		unlock_page(page);
356 	return ret;
357 }
358 
359 /*
360  * This is used only for read-ahead. Failures or difficult to handle
361  * situations are safe to ignore.
362  *
363  * Right now, we don't bother with BH_Boundary - in-inode extent lists
364  * are quite large (243 extents on 4k blocks), so most inodes don't
365  * grow out to a tree. If need be, detecting boundary extents could
366  * trivially be added in a future version of ocfs2_get_block().
367  */
368 static int ocfs2_readpages(struct file *filp, struct address_space *mapping,
369 			   struct list_head *pages, unsigned nr_pages)
370 {
371 	int ret, err = -EIO;
372 	struct inode *inode = mapping->host;
373 	struct ocfs2_inode_info *oi = OCFS2_I(inode);
374 	loff_t start;
375 	struct page *last;
376 
377 	/*
378 	 * Use the nonblocking flag for the dlm code to avoid page
379 	 * lock inversion, but don't bother with retrying.
380 	 */
381 	ret = ocfs2_inode_lock_full(inode, NULL, 0, OCFS2_LOCK_NONBLOCK);
382 	if (ret)
383 		return err;
384 
385 	if (down_read_trylock(&oi->ip_alloc_sem) == 0) {
386 		ocfs2_inode_unlock(inode, 0);
387 		return err;
388 	}
389 
390 	/*
391 	 * Don't bother with inline-data. There isn't anything
392 	 * to read-ahead in that case anyway...
393 	 */
394 	if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL)
395 		goto out_unlock;
396 
397 	/*
398 	 * Check whether a remote node truncated this file - we just
399 	 * drop out in that case as it's not worth handling here.
400 	 */
401 	last = lru_to_page(pages);
402 	start = (loff_t)last->index << PAGE_SHIFT;
403 	if (start >= i_size_read(inode))
404 		goto out_unlock;
405 
406 	err = mpage_readpages(mapping, pages, nr_pages, ocfs2_get_block);
407 
408 out_unlock:
409 	up_read(&oi->ip_alloc_sem);
410 	ocfs2_inode_unlock(inode, 0);
411 
412 	return err;
413 }
414 
415 /* Note: Because we don't support holes, our allocation has
416  * already happened (allocation writes zeros to the file data)
417  * so we don't have to worry about ordered writes in
418  * ocfs2_writepage.
419  *
420  * ->writepage is called during the process of invalidating the page cache
421  * during blocked lock processing.  It can't block on any cluster locks
422  * to during block mapping.  It's relying on the fact that the block
423  * mapping can't have disappeared under the dirty pages that it is
424  * being asked to write back.
425  */
426 static int ocfs2_writepage(struct page *page, struct writeback_control *wbc)
427 {
428 	trace_ocfs2_writepage(
429 		(unsigned long long)OCFS2_I(page->mapping->host)->ip_blkno,
430 		page->index);
431 
432 	return block_write_full_page(page, ocfs2_get_block, wbc);
433 }
434 
435 /* Taken from ext3. We don't necessarily need the full blown
436  * functionality yet, but IMHO it's better to cut and paste the whole
437  * thing so we can avoid introducing our own bugs (and easily pick up
438  * their fixes when they happen) --Mark */
439 int walk_page_buffers(	handle_t *handle,
440 			struct buffer_head *head,
441 			unsigned from,
442 			unsigned to,
443 			int *partial,
444 			int (*fn)(	handle_t *handle,
445 					struct buffer_head *bh))
446 {
447 	struct buffer_head *bh;
448 	unsigned block_start, block_end;
449 	unsigned blocksize = head->b_size;
450 	int err, ret = 0;
451 	struct buffer_head *next;
452 
453 	for (	bh = head, block_start = 0;
454 		ret == 0 && (bh != head || !block_start);
455 	    	block_start = block_end, bh = next)
456 	{
457 		next = bh->b_this_page;
458 		block_end = block_start + blocksize;
459 		if (block_end <= from || block_start >= to) {
460 			if (partial && !buffer_uptodate(bh))
461 				*partial = 1;
462 			continue;
463 		}
464 		err = (*fn)(handle, bh);
465 		if (!ret)
466 			ret = err;
467 	}
468 	return ret;
469 }
470 
471 static sector_t ocfs2_bmap(struct address_space *mapping, sector_t block)
472 {
473 	sector_t status;
474 	u64 p_blkno = 0;
475 	int err = 0;
476 	struct inode *inode = mapping->host;
477 
478 	trace_ocfs2_bmap((unsigned long long)OCFS2_I(inode)->ip_blkno,
479 			 (unsigned long long)block);
480 
481 	/*
482 	 * The swap code (ab-)uses ->bmap to get a block mapping and then
483 	 * bypasseѕ the file system for actual I/O.  We really can't allow
484 	 * that on refcounted inodes, so we have to skip out here.  And yes,
485 	 * 0 is the magic code for a bmap error..
486 	 */
487 	if (ocfs2_is_refcount_inode(inode))
488 		return 0;
489 
490 	/* We don't need to lock journal system files, since they aren't
491 	 * accessed concurrently from multiple nodes.
492 	 */
493 	if (!INODE_JOURNAL(inode)) {
494 		err = ocfs2_inode_lock(inode, NULL, 0);
495 		if (err) {
496 			if (err != -ENOENT)
497 				mlog_errno(err);
498 			goto bail;
499 		}
500 		down_read(&OCFS2_I(inode)->ip_alloc_sem);
501 	}
502 
503 	if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL))
504 		err = ocfs2_extent_map_get_blocks(inode, block, &p_blkno, NULL,
505 						  NULL);
506 
507 	if (!INODE_JOURNAL(inode)) {
508 		up_read(&OCFS2_I(inode)->ip_alloc_sem);
509 		ocfs2_inode_unlock(inode, 0);
510 	}
511 
512 	if (err) {
513 		mlog(ML_ERROR, "get_blocks() failed, block = %llu\n",
514 		     (unsigned long long)block);
515 		mlog_errno(err);
516 		goto bail;
517 	}
518 
519 bail:
520 	status = err ? 0 : p_blkno;
521 
522 	return status;
523 }
524 
525 static int ocfs2_releasepage(struct page *page, gfp_t wait)
526 {
527 	if (!page_has_buffers(page))
528 		return 0;
529 	return try_to_free_buffers(page);
530 }
531 
532 static void ocfs2_figure_cluster_boundaries(struct ocfs2_super *osb,
533 					    u32 cpos,
534 					    unsigned int *start,
535 					    unsigned int *end)
536 {
537 	unsigned int cluster_start = 0, cluster_end = PAGE_SIZE;
538 
539 	if (unlikely(PAGE_SHIFT > osb->s_clustersize_bits)) {
540 		unsigned int cpp;
541 
542 		cpp = 1 << (PAGE_SHIFT - osb->s_clustersize_bits);
543 
544 		cluster_start = cpos % cpp;
545 		cluster_start = cluster_start << osb->s_clustersize_bits;
546 
547 		cluster_end = cluster_start + osb->s_clustersize;
548 	}
549 
550 	BUG_ON(cluster_start > PAGE_SIZE);
551 	BUG_ON(cluster_end > PAGE_SIZE);
552 
553 	if (start)
554 		*start = cluster_start;
555 	if (end)
556 		*end = cluster_end;
557 }
558 
559 /*
560  * 'from' and 'to' are the region in the page to avoid zeroing.
561  *
562  * If pagesize > clustersize, this function will avoid zeroing outside
563  * of the cluster boundary.
564  *
565  * from == to == 0 is code for "zero the entire cluster region"
566  */
567 static void ocfs2_clear_page_regions(struct page *page,
568 				     struct ocfs2_super *osb, u32 cpos,
569 				     unsigned from, unsigned to)
570 {
571 	void *kaddr;
572 	unsigned int cluster_start, cluster_end;
573 
574 	ocfs2_figure_cluster_boundaries(osb, cpos, &cluster_start, &cluster_end);
575 
576 	kaddr = kmap_atomic(page);
577 
578 	if (from || to) {
579 		if (from > cluster_start)
580 			memset(kaddr + cluster_start, 0, from - cluster_start);
581 		if (to < cluster_end)
582 			memset(kaddr + to, 0, cluster_end - to);
583 	} else {
584 		memset(kaddr + cluster_start, 0, cluster_end - cluster_start);
585 	}
586 
587 	kunmap_atomic(kaddr);
588 }
589 
590 /*
591  * Nonsparse file systems fully allocate before we get to the write
592  * code. This prevents ocfs2_write() from tagging the write as an
593  * allocating one, which means ocfs2_map_page_blocks() might try to
594  * read-in the blocks at the tail of our file. Avoid reading them by
595  * testing i_size against each block offset.
596  */
597 static int ocfs2_should_read_blk(struct inode *inode, struct page *page,
598 				 unsigned int block_start)
599 {
600 	u64 offset = page_offset(page) + block_start;
601 
602 	if (ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)))
603 		return 1;
604 
605 	if (i_size_read(inode) > offset)
606 		return 1;
607 
608 	return 0;
609 }
610 
611 /*
612  * Some of this taken from __block_write_begin(). We already have our
613  * mapping by now though, and the entire write will be allocating or
614  * it won't, so not much need to use BH_New.
615  *
616  * This will also skip zeroing, which is handled externally.
617  */
618 int ocfs2_map_page_blocks(struct page *page, u64 *p_blkno,
619 			  struct inode *inode, unsigned int from,
620 			  unsigned int to, int new)
621 {
622 	int ret = 0;
623 	struct buffer_head *head, *bh, *wait[2], **wait_bh = wait;
624 	unsigned int block_end, block_start;
625 	unsigned int bsize = i_blocksize(inode);
626 
627 	if (!page_has_buffers(page))
628 		create_empty_buffers(page, bsize, 0);
629 
630 	head = page_buffers(page);
631 	for (bh = head, block_start = 0; bh != head || !block_start;
632 	     bh = bh->b_this_page, block_start += bsize) {
633 		block_end = block_start + bsize;
634 
635 		clear_buffer_new(bh);
636 
637 		/*
638 		 * Ignore blocks outside of our i/o range -
639 		 * they may belong to unallocated clusters.
640 		 */
641 		if (block_start >= to || block_end <= from) {
642 			if (PageUptodate(page))
643 				set_buffer_uptodate(bh);
644 			continue;
645 		}
646 
647 		/*
648 		 * For an allocating write with cluster size >= page
649 		 * size, we always write the entire page.
650 		 */
651 		if (new)
652 			set_buffer_new(bh);
653 
654 		if (!buffer_mapped(bh)) {
655 			map_bh(bh, inode->i_sb, *p_blkno);
656 			clean_bdev_bh_alias(bh);
657 		}
658 
659 		if (PageUptodate(page)) {
660 			if (!buffer_uptodate(bh))
661 				set_buffer_uptodate(bh);
662 		} else if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
663 			   !buffer_new(bh) &&
664 			   ocfs2_should_read_blk(inode, page, block_start) &&
665 			   (block_start < from || block_end > to)) {
666 			ll_rw_block(REQ_OP_READ, 0, 1, &bh);
667 			*wait_bh++=bh;
668 		}
669 
670 		*p_blkno = *p_blkno + 1;
671 	}
672 
673 	/*
674 	 * If we issued read requests - let them complete.
675 	 */
676 	while(wait_bh > wait) {
677 		wait_on_buffer(*--wait_bh);
678 		if (!buffer_uptodate(*wait_bh))
679 			ret = -EIO;
680 	}
681 
682 	if (ret == 0 || !new)
683 		return ret;
684 
685 	/*
686 	 * If we get -EIO above, zero out any newly allocated blocks
687 	 * to avoid exposing stale data.
688 	 */
689 	bh = head;
690 	block_start = 0;
691 	do {
692 		block_end = block_start + bsize;
693 		if (block_end <= from)
694 			goto next_bh;
695 		if (block_start >= to)
696 			break;
697 
698 		zero_user(page, block_start, bh->b_size);
699 		set_buffer_uptodate(bh);
700 		mark_buffer_dirty(bh);
701 
702 next_bh:
703 		block_start = block_end;
704 		bh = bh->b_this_page;
705 	} while (bh != head);
706 
707 	return ret;
708 }
709 
710 #if (PAGE_SIZE >= OCFS2_MAX_CLUSTERSIZE)
711 #define OCFS2_MAX_CTXT_PAGES	1
712 #else
713 #define OCFS2_MAX_CTXT_PAGES	(OCFS2_MAX_CLUSTERSIZE / PAGE_SIZE)
714 #endif
715 
716 #define OCFS2_MAX_CLUSTERS_PER_PAGE	(PAGE_SIZE / OCFS2_MIN_CLUSTERSIZE)
717 
718 struct ocfs2_unwritten_extent {
719 	struct list_head	ue_node;
720 	struct list_head	ue_ip_node;
721 	u32			ue_cpos;
722 	u32			ue_phys;
723 };
724 
725 /*
726  * Describe the state of a single cluster to be written to.
727  */
728 struct ocfs2_write_cluster_desc {
729 	u32		c_cpos;
730 	u32		c_phys;
731 	/*
732 	 * Give this a unique field because c_phys eventually gets
733 	 * filled.
734 	 */
735 	unsigned	c_new;
736 	unsigned	c_clear_unwritten;
737 	unsigned	c_needs_zero;
738 };
739 
740 struct ocfs2_write_ctxt {
741 	/* Logical cluster position / len of write */
742 	u32				w_cpos;
743 	u32				w_clen;
744 
745 	/* First cluster allocated in a nonsparse extend */
746 	u32				w_first_new_cpos;
747 
748 	/* Type of caller. Must be one of buffer, mmap, direct.  */
749 	ocfs2_write_type_t		w_type;
750 
751 	struct ocfs2_write_cluster_desc	w_desc[OCFS2_MAX_CLUSTERS_PER_PAGE];
752 
753 	/*
754 	 * This is true if page_size > cluster_size.
755 	 *
756 	 * It triggers a set of special cases during write which might
757 	 * have to deal with allocating writes to partial pages.
758 	 */
759 	unsigned int			w_large_pages;
760 
761 	/*
762 	 * Pages involved in this write.
763 	 *
764 	 * w_target_page is the page being written to by the user.
765 	 *
766 	 * w_pages is an array of pages which always contains
767 	 * w_target_page, and in the case of an allocating write with
768 	 * page_size < cluster size, it will contain zero'd and mapped
769 	 * pages adjacent to w_target_page which need to be written
770 	 * out in so that future reads from that region will get
771 	 * zero's.
772 	 */
773 	unsigned int			w_num_pages;
774 	struct page			*w_pages[OCFS2_MAX_CTXT_PAGES];
775 	struct page			*w_target_page;
776 
777 	/*
778 	 * w_target_locked is used for page_mkwrite path indicating no unlocking
779 	 * against w_target_page in ocfs2_write_end_nolock.
780 	 */
781 	unsigned int			w_target_locked:1;
782 
783 	/*
784 	 * ocfs2_write_end() uses this to know what the real range to
785 	 * write in the target should be.
786 	 */
787 	unsigned int			w_target_from;
788 	unsigned int			w_target_to;
789 
790 	/*
791 	 * We could use journal_current_handle() but this is cleaner,
792 	 * IMHO -Mark
793 	 */
794 	handle_t			*w_handle;
795 
796 	struct buffer_head		*w_di_bh;
797 
798 	struct ocfs2_cached_dealloc_ctxt w_dealloc;
799 
800 	struct list_head		w_unwritten_list;
801 	unsigned int			w_unwritten_count;
802 };
803 
804 void ocfs2_unlock_and_free_pages(struct page **pages, int num_pages)
805 {
806 	int i;
807 
808 	for(i = 0; i < num_pages; i++) {
809 		if (pages[i]) {
810 			unlock_page(pages[i]);
811 			mark_page_accessed(pages[i]);
812 			put_page(pages[i]);
813 		}
814 	}
815 }
816 
817 static void ocfs2_unlock_pages(struct ocfs2_write_ctxt *wc)
818 {
819 	int i;
820 
821 	/*
822 	 * w_target_locked is only set to true in the page_mkwrite() case.
823 	 * The intent is to allow us to lock the target page from write_begin()
824 	 * to write_end(). The caller must hold a ref on w_target_page.
825 	 */
826 	if (wc->w_target_locked) {
827 		BUG_ON(!wc->w_target_page);
828 		for (i = 0; i < wc->w_num_pages; i++) {
829 			if (wc->w_target_page == wc->w_pages[i]) {
830 				wc->w_pages[i] = NULL;
831 				break;
832 			}
833 		}
834 		mark_page_accessed(wc->w_target_page);
835 		put_page(wc->w_target_page);
836 	}
837 	ocfs2_unlock_and_free_pages(wc->w_pages, wc->w_num_pages);
838 }
839 
840 static void ocfs2_free_unwritten_list(struct inode *inode,
841 				 struct list_head *head)
842 {
843 	struct ocfs2_inode_info *oi = OCFS2_I(inode);
844 	struct ocfs2_unwritten_extent *ue = NULL, *tmp = NULL;
845 
846 	list_for_each_entry_safe(ue, tmp, head, ue_node) {
847 		list_del(&ue->ue_node);
848 		spin_lock(&oi->ip_lock);
849 		list_del(&ue->ue_ip_node);
850 		spin_unlock(&oi->ip_lock);
851 		kfree(ue);
852 	}
853 }
854 
855 static void ocfs2_free_write_ctxt(struct inode *inode,
856 				  struct ocfs2_write_ctxt *wc)
857 {
858 	ocfs2_free_unwritten_list(inode, &wc->w_unwritten_list);
859 	ocfs2_unlock_pages(wc);
860 	brelse(wc->w_di_bh);
861 	kfree(wc);
862 }
863 
864 static int ocfs2_alloc_write_ctxt(struct ocfs2_write_ctxt **wcp,
865 				  struct ocfs2_super *osb, loff_t pos,
866 				  unsigned len, ocfs2_write_type_t type,
867 				  struct buffer_head *di_bh)
868 {
869 	u32 cend;
870 	struct ocfs2_write_ctxt *wc;
871 
872 	wc = kzalloc(sizeof(struct ocfs2_write_ctxt), GFP_NOFS);
873 	if (!wc)
874 		return -ENOMEM;
875 
876 	wc->w_cpos = pos >> osb->s_clustersize_bits;
877 	wc->w_first_new_cpos = UINT_MAX;
878 	cend = (pos + len - 1) >> osb->s_clustersize_bits;
879 	wc->w_clen = cend - wc->w_cpos + 1;
880 	get_bh(di_bh);
881 	wc->w_di_bh = di_bh;
882 	wc->w_type = type;
883 
884 	if (unlikely(PAGE_SHIFT > osb->s_clustersize_bits))
885 		wc->w_large_pages = 1;
886 	else
887 		wc->w_large_pages = 0;
888 
889 	ocfs2_init_dealloc_ctxt(&wc->w_dealloc);
890 	INIT_LIST_HEAD(&wc->w_unwritten_list);
891 
892 	*wcp = wc;
893 
894 	return 0;
895 }
896 
897 /*
898  * If a page has any new buffers, zero them out here, and mark them uptodate
899  * and dirty so they'll be written out (in order to prevent uninitialised
900  * block data from leaking). And clear the new bit.
901  */
902 static void ocfs2_zero_new_buffers(struct page *page, unsigned from, unsigned to)
903 {
904 	unsigned int block_start, block_end;
905 	struct buffer_head *head, *bh;
906 
907 	BUG_ON(!PageLocked(page));
908 	if (!page_has_buffers(page))
909 		return;
910 
911 	bh = head = page_buffers(page);
912 	block_start = 0;
913 	do {
914 		block_end = block_start + bh->b_size;
915 
916 		if (buffer_new(bh)) {
917 			if (block_end > from && block_start < to) {
918 				if (!PageUptodate(page)) {
919 					unsigned start, end;
920 
921 					start = max(from, block_start);
922 					end = min(to, block_end);
923 
924 					zero_user_segment(page, start, end);
925 					set_buffer_uptodate(bh);
926 				}
927 
928 				clear_buffer_new(bh);
929 				mark_buffer_dirty(bh);
930 			}
931 		}
932 
933 		block_start = block_end;
934 		bh = bh->b_this_page;
935 	} while (bh != head);
936 }
937 
938 /*
939  * Only called when we have a failure during allocating write to write
940  * zero's to the newly allocated region.
941  */
942 static void ocfs2_write_failure(struct inode *inode,
943 				struct ocfs2_write_ctxt *wc,
944 				loff_t user_pos, unsigned user_len)
945 {
946 	int i;
947 	unsigned from = user_pos & (PAGE_SIZE - 1),
948 		to = user_pos + user_len;
949 	struct page *tmppage;
950 
951 	if (wc->w_target_page)
952 		ocfs2_zero_new_buffers(wc->w_target_page, from, to);
953 
954 	for(i = 0; i < wc->w_num_pages; i++) {
955 		tmppage = wc->w_pages[i];
956 
957 		if (tmppage && page_has_buffers(tmppage)) {
958 			if (ocfs2_should_order_data(inode))
959 				ocfs2_jbd2_file_inode(wc->w_handle, inode);
960 
961 			block_commit_write(tmppage, from, to);
962 		}
963 	}
964 }
965 
966 static int ocfs2_prepare_page_for_write(struct inode *inode, u64 *p_blkno,
967 					struct ocfs2_write_ctxt *wc,
968 					struct page *page, u32 cpos,
969 					loff_t user_pos, unsigned user_len,
970 					int new)
971 {
972 	int ret;
973 	unsigned int map_from = 0, map_to = 0;
974 	unsigned int cluster_start, cluster_end;
975 	unsigned int user_data_from = 0, user_data_to = 0;
976 
977 	ocfs2_figure_cluster_boundaries(OCFS2_SB(inode->i_sb), cpos,
978 					&cluster_start, &cluster_end);
979 
980 	/* treat the write as new if the a hole/lseek spanned across
981 	 * the page boundary.
982 	 */
983 	new = new | ((i_size_read(inode) <= page_offset(page)) &&
984 			(page_offset(page) <= user_pos));
985 
986 	if (page == wc->w_target_page) {
987 		map_from = user_pos & (PAGE_SIZE - 1);
988 		map_to = map_from + user_len;
989 
990 		if (new)
991 			ret = ocfs2_map_page_blocks(page, p_blkno, inode,
992 						    cluster_start, cluster_end,
993 						    new);
994 		else
995 			ret = ocfs2_map_page_blocks(page, p_blkno, inode,
996 						    map_from, map_to, new);
997 		if (ret) {
998 			mlog_errno(ret);
999 			goto out;
1000 		}
1001 
1002 		user_data_from = map_from;
1003 		user_data_to = map_to;
1004 		if (new) {
1005 			map_from = cluster_start;
1006 			map_to = cluster_end;
1007 		}
1008 	} else {
1009 		/*
1010 		 * If we haven't allocated the new page yet, we
1011 		 * shouldn't be writing it out without copying user
1012 		 * data. This is likely a math error from the caller.
1013 		 */
1014 		BUG_ON(!new);
1015 
1016 		map_from = cluster_start;
1017 		map_to = cluster_end;
1018 
1019 		ret = ocfs2_map_page_blocks(page, p_blkno, inode,
1020 					    cluster_start, cluster_end, new);
1021 		if (ret) {
1022 			mlog_errno(ret);
1023 			goto out;
1024 		}
1025 	}
1026 
1027 	/*
1028 	 * Parts of newly allocated pages need to be zero'd.
1029 	 *
1030 	 * Above, we have also rewritten 'to' and 'from' - as far as
1031 	 * the rest of the function is concerned, the entire cluster
1032 	 * range inside of a page needs to be written.
1033 	 *
1034 	 * We can skip this if the page is up to date - it's already
1035 	 * been zero'd from being read in as a hole.
1036 	 */
1037 	if (new && !PageUptodate(page))
1038 		ocfs2_clear_page_regions(page, OCFS2_SB(inode->i_sb),
1039 					 cpos, user_data_from, user_data_to);
1040 
1041 	flush_dcache_page(page);
1042 
1043 out:
1044 	return ret;
1045 }
1046 
1047 /*
1048  * This function will only grab one clusters worth of pages.
1049  */
1050 static int ocfs2_grab_pages_for_write(struct address_space *mapping,
1051 				      struct ocfs2_write_ctxt *wc,
1052 				      u32 cpos, loff_t user_pos,
1053 				      unsigned user_len, int new,
1054 				      struct page *mmap_page)
1055 {
1056 	int ret = 0, i;
1057 	unsigned long start, target_index, end_index, index;
1058 	struct inode *inode = mapping->host;
1059 	loff_t last_byte;
1060 
1061 	target_index = user_pos >> PAGE_SHIFT;
1062 
1063 	/*
1064 	 * Figure out how many pages we'll be manipulating here. For
1065 	 * non allocating write, we just change the one
1066 	 * page. Otherwise, we'll need a whole clusters worth.  If we're
1067 	 * writing past i_size, we only need enough pages to cover the
1068 	 * last page of the write.
1069 	 */
1070 	if (new) {
1071 		wc->w_num_pages = ocfs2_pages_per_cluster(inode->i_sb);
1072 		start = ocfs2_align_clusters_to_page_index(inode->i_sb, cpos);
1073 		/*
1074 		 * We need the index *past* the last page we could possibly
1075 		 * touch.  This is the page past the end of the write or
1076 		 * i_size, whichever is greater.
1077 		 */
1078 		last_byte = max(user_pos + user_len, i_size_read(inode));
1079 		BUG_ON(last_byte < 1);
1080 		end_index = ((last_byte - 1) >> PAGE_SHIFT) + 1;
1081 		if ((start + wc->w_num_pages) > end_index)
1082 			wc->w_num_pages = end_index - start;
1083 	} else {
1084 		wc->w_num_pages = 1;
1085 		start = target_index;
1086 	}
1087 	end_index = (user_pos + user_len - 1) >> PAGE_SHIFT;
1088 
1089 	for(i = 0; i < wc->w_num_pages; i++) {
1090 		index = start + i;
1091 
1092 		if (index >= target_index && index <= end_index &&
1093 		    wc->w_type == OCFS2_WRITE_MMAP) {
1094 			/*
1095 			 * ocfs2_pagemkwrite() is a little different
1096 			 * and wants us to directly use the page
1097 			 * passed in.
1098 			 */
1099 			lock_page(mmap_page);
1100 
1101 			/* Exit and let the caller retry */
1102 			if (mmap_page->mapping != mapping) {
1103 				WARN_ON(mmap_page->mapping);
1104 				unlock_page(mmap_page);
1105 				ret = -EAGAIN;
1106 				goto out;
1107 			}
1108 
1109 			get_page(mmap_page);
1110 			wc->w_pages[i] = mmap_page;
1111 			wc->w_target_locked = true;
1112 		} else if (index >= target_index && index <= end_index &&
1113 			   wc->w_type == OCFS2_WRITE_DIRECT) {
1114 			/* Direct write has no mapping page. */
1115 			wc->w_pages[i] = NULL;
1116 			continue;
1117 		} else {
1118 			wc->w_pages[i] = find_or_create_page(mapping, index,
1119 							     GFP_NOFS);
1120 			if (!wc->w_pages[i]) {
1121 				ret = -ENOMEM;
1122 				mlog_errno(ret);
1123 				goto out;
1124 			}
1125 		}
1126 		wait_for_stable_page(wc->w_pages[i]);
1127 
1128 		if (index == target_index)
1129 			wc->w_target_page = wc->w_pages[i];
1130 	}
1131 out:
1132 	if (ret)
1133 		wc->w_target_locked = false;
1134 	return ret;
1135 }
1136 
1137 /*
1138  * Prepare a single cluster for write one cluster into the file.
1139  */
1140 static int ocfs2_write_cluster(struct address_space *mapping,
1141 			       u32 *phys, unsigned int new,
1142 			       unsigned int clear_unwritten,
1143 			       unsigned int should_zero,
1144 			       struct ocfs2_alloc_context *data_ac,
1145 			       struct ocfs2_alloc_context *meta_ac,
1146 			       struct ocfs2_write_ctxt *wc, u32 cpos,
1147 			       loff_t user_pos, unsigned user_len)
1148 {
1149 	int ret, i;
1150 	u64 p_blkno;
1151 	struct inode *inode = mapping->host;
1152 	struct ocfs2_extent_tree et;
1153 	int bpc = ocfs2_clusters_to_blocks(inode->i_sb, 1);
1154 
1155 	if (new) {
1156 		u32 tmp_pos;
1157 
1158 		/*
1159 		 * This is safe to call with the page locks - it won't take
1160 		 * any additional semaphores or cluster locks.
1161 		 */
1162 		tmp_pos = cpos;
1163 		ret = ocfs2_add_inode_data(OCFS2_SB(inode->i_sb), inode,
1164 					   &tmp_pos, 1, !clear_unwritten,
1165 					   wc->w_di_bh, wc->w_handle,
1166 					   data_ac, meta_ac, NULL);
1167 		/*
1168 		 * This shouldn't happen because we must have already
1169 		 * calculated the correct meta data allocation required. The
1170 		 * internal tree allocation code should know how to increase
1171 		 * transaction credits itself.
1172 		 *
1173 		 * If need be, we could handle -EAGAIN for a
1174 		 * RESTART_TRANS here.
1175 		 */
1176 		mlog_bug_on_msg(ret == -EAGAIN,
1177 				"Inode %llu: EAGAIN return during allocation.\n",
1178 				(unsigned long long)OCFS2_I(inode)->ip_blkno);
1179 		if (ret < 0) {
1180 			mlog_errno(ret);
1181 			goto out;
1182 		}
1183 	} else if (clear_unwritten) {
1184 		ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode),
1185 					      wc->w_di_bh);
1186 		ret = ocfs2_mark_extent_written(inode, &et,
1187 						wc->w_handle, cpos, 1, *phys,
1188 						meta_ac, &wc->w_dealloc);
1189 		if (ret < 0) {
1190 			mlog_errno(ret);
1191 			goto out;
1192 		}
1193 	}
1194 
1195 	/*
1196 	 * The only reason this should fail is due to an inability to
1197 	 * find the extent added.
1198 	 */
1199 	ret = ocfs2_get_clusters(inode, cpos, phys, NULL, NULL);
1200 	if (ret < 0) {
1201 		mlog(ML_ERROR, "Get physical blkno failed for inode %llu, "
1202 			    "at logical cluster %u",
1203 			    (unsigned long long)OCFS2_I(inode)->ip_blkno, cpos);
1204 		goto out;
1205 	}
1206 
1207 	BUG_ON(*phys == 0);
1208 
1209 	p_blkno = ocfs2_clusters_to_blocks(inode->i_sb, *phys);
1210 	if (!should_zero)
1211 		p_blkno += (user_pos >> inode->i_sb->s_blocksize_bits) & (u64)(bpc - 1);
1212 
1213 	for(i = 0; i < wc->w_num_pages; i++) {
1214 		int tmpret;
1215 
1216 		/* This is the direct io target page. */
1217 		if (wc->w_pages[i] == NULL) {
1218 			p_blkno++;
1219 			continue;
1220 		}
1221 
1222 		tmpret = ocfs2_prepare_page_for_write(inode, &p_blkno, wc,
1223 						      wc->w_pages[i], cpos,
1224 						      user_pos, user_len,
1225 						      should_zero);
1226 		if (tmpret) {
1227 			mlog_errno(tmpret);
1228 			if (ret == 0)
1229 				ret = tmpret;
1230 		}
1231 	}
1232 
1233 	/*
1234 	 * We only have cleanup to do in case of allocating write.
1235 	 */
1236 	if (ret && new)
1237 		ocfs2_write_failure(inode, wc, user_pos, user_len);
1238 
1239 out:
1240 
1241 	return ret;
1242 }
1243 
1244 static int ocfs2_write_cluster_by_desc(struct address_space *mapping,
1245 				       struct ocfs2_alloc_context *data_ac,
1246 				       struct ocfs2_alloc_context *meta_ac,
1247 				       struct ocfs2_write_ctxt *wc,
1248 				       loff_t pos, unsigned len)
1249 {
1250 	int ret, i;
1251 	loff_t cluster_off;
1252 	unsigned int local_len = len;
1253 	struct ocfs2_write_cluster_desc *desc;
1254 	struct ocfs2_super *osb = OCFS2_SB(mapping->host->i_sb);
1255 
1256 	for (i = 0; i < wc->w_clen; i++) {
1257 		desc = &wc->w_desc[i];
1258 
1259 		/*
1260 		 * We have to make sure that the total write passed in
1261 		 * doesn't extend past a single cluster.
1262 		 */
1263 		local_len = len;
1264 		cluster_off = pos & (osb->s_clustersize - 1);
1265 		if ((cluster_off + local_len) > osb->s_clustersize)
1266 			local_len = osb->s_clustersize - cluster_off;
1267 
1268 		ret = ocfs2_write_cluster(mapping, &desc->c_phys,
1269 					  desc->c_new,
1270 					  desc->c_clear_unwritten,
1271 					  desc->c_needs_zero,
1272 					  data_ac, meta_ac,
1273 					  wc, desc->c_cpos, pos, local_len);
1274 		if (ret) {
1275 			mlog_errno(ret);
1276 			goto out;
1277 		}
1278 
1279 		len -= local_len;
1280 		pos += local_len;
1281 	}
1282 
1283 	ret = 0;
1284 out:
1285 	return ret;
1286 }
1287 
1288 /*
1289  * ocfs2_write_end() wants to know which parts of the target page it
1290  * should complete the write on. It's easiest to compute them ahead of
1291  * time when a more complete view of the write is available.
1292  */
1293 static void ocfs2_set_target_boundaries(struct ocfs2_super *osb,
1294 					struct ocfs2_write_ctxt *wc,
1295 					loff_t pos, unsigned len, int alloc)
1296 {
1297 	struct ocfs2_write_cluster_desc *desc;
1298 
1299 	wc->w_target_from = pos & (PAGE_SIZE - 1);
1300 	wc->w_target_to = wc->w_target_from + len;
1301 
1302 	if (alloc == 0)
1303 		return;
1304 
1305 	/*
1306 	 * Allocating write - we may have different boundaries based
1307 	 * on page size and cluster size.
1308 	 *
1309 	 * NOTE: We can no longer compute one value from the other as
1310 	 * the actual write length and user provided length may be
1311 	 * different.
1312 	 */
1313 
1314 	if (wc->w_large_pages) {
1315 		/*
1316 		 * We only care about the 1st and last cluster within
1317 		 * our range and whether they should be zero'd or not. Either
1318 		 * value may be extended out to the start/end of a
1319 		 * newly allocated cluster.
1320 		 */
1321 		desc = &wc->w_desc[0];
1322 		if (desc->c_needs_zero)
1323 			ocfs2_figure_cluster_boundaries(osb,
1324 							desc->c_cpos,
1325 							&wc->w_target_from,
1326 							NULL);
1327 
1328 		desc = &wc->w_desc[wc->w_clen - 1];
1329 		if (desc->c_needs_zero)
1330 			ocfs2_figure_cluster_boundaries(osb,
1331 							desc->c_cpos,
1332 							NULL,
1333 							&wc->w_target_to);
1334 	} else {
1335 		wc->w_target_from = 0;
1336 		wc->w_target_to = PAGE_SIZE;
1337 	}
1338 }
1339 
1340 /*
1341  * Check if this extent is marked UNWRITTEN by direct io. If so, we need not to
1342  * do the zero work. And should not to clear UNWRITTEN since it will be cleared
1343  * by the direct io procedure.
1344  * If this is a new extent that allocated by direct io, we should mark it in
1345  * the ip_unwritten_list.
1346  */
1347 static int ocfs2_unwritten_check(struct inode *inode,
1348 				 struct ocfs2_write_ctxt *wc,
1349 				 struct ocfs2_write_cluster_desc *desc)
1350 {
1351 	struct ocfs2_inode_info *oi = OCFS2_I(inode);
1352 	struct ocfs2_unwritten_extent *ue = NULL, *new = NULL;
1353 	int ret = 0;
1354 
1355 	if (!desc->c_needs_zero)
1356 		return 0;
1357 
1358 retry:
1359 	spin_lock(&oi->ip_lock);
1360 	/* Needs not to zero no metter buffer or direct. The one who is zero
1361 	 * the cluster is doing zero. And he will clear unwritten after all
1362 	 * cluster io finished. */
1363 	list_for_each_entry(ue, &oi->ip_unwritten_list, ue_ip_node) {
1364 		if (desc->c_cpos == ue->ue_cpos) {
1365 			BUG_ON(desc->c_new);
1366 			desc->c_needs_zero = 0;
1367 			desc->c_clear_unwritten = 0;
1368 			goto unlock;
1369 		}
1370 	}
1371 
1372 	if (wc->w_type != OCFS2_WRITE_DIRECT)
1373 		goto unlock;
1374 
1375 	if (new == NULL) {
1376 		spin_unlock(&oi->ip_lock);
1377 		new = kmalloc(sizeof(struct ocfs2_unwritten_extent),
1378 			     GFP_NOFS);
1379 		if (new == NULL) {
1380 			ret = -ENOMEM;
1381 			goto out;
1382 		}
1383 		goto retry;
1384 	}
1385 	/* This direct write will doing zero. */
1386 	new->ue_cpos = desc->c_cpos;
1387 	new->ue_phys = desc->c_phys;
1388 	desc->c_clear_unwritten = 0;
1389 	list_add_tail(&new->ue_ip_node, &oi->ip_unwritten_list);
1390 	list_add_tail(&new->ue_node, &wc->w_unwritten_list);
1391 	wc->w_unwritten_count++;
1392 	new = NULL;
1393 unlock:
1394 	spin_unlock(&oi->ip_lock);
1395 out:
1396 	kfree(new);
1397 	return ret;
1398 }
1399 
1400 /*
1401  * Populate each single-cluster write descriptor in the write context
1402  * with information about the i/o to be done.
1403  *
1404  * Returns the number of clusters that will have to be allocated, as
1405  * well as a worst case estimate of the number of extent records that
1406  * would have to be created during a write to an unwritten region.
1407  */
1408 static int ocfs2_populate_write_desc(struct inode *inode,
1409 				     struct ocfs2_write_ctxt *wc,
1410 				     unsigned int *clusters_to_alloc,
1411 				     unsigned int *extents_to_split)
1412 {
1413 	int ret;
1414 	struct ocfs2_write_cluster_desc *desc;
1415 	unsigned int num_clusters = 0;
1416 	unsigned int ext_flags = 0;
1417 	u32 phys = 0;
1418 	int i;
1419 
1420 	*clusters_to_alloc = 0;
1421 	*extents_to_split = 0;
1422 
1423 	for (i = 0; i < wc->w_clen; i++) {
1424 		desc = &wc->w_desc[i];
1425 		desc->c_cpos = wc->w_cpos + i;
1426 
1427 		if (num_clusters == 0) {
1428 			/*
1429 			 * Need to look up the next extent record.
1430 			 */
1431 			ret = ocfs2_get_clusters(inode, desc->c_cpos, &phys,
1432 						 &num_clusters, &ext_flags);
1433 			if (ret) {
1434 				mlog_errno(ret);
1435 				goto out;
1436 			}
1437 
1438 			/* We should already CoW the refcountd extent. */
1439 			BUG_ON(ext_flags & OCFS2_EXT_REFCOUNTED);
1440 
1441 			/*
1442 			 * Assume worst case - that we're writing in
1443 			 * the middle of the extent.
1444 			 *
1445 			 * We can assume that the write proceeds from
1446 			 * left to right, in which case the extent
1447 			 * insert code is smart enough to coalesce the
1448 			 * next splits into the previous records created.
1449 			 */
1450 			if (ext_flags & OCFS2_EXT_UNWRITTEN)
1451 				*extents_to_split = *extents_to_split + 2;
1452 		} else if (phys) {
1453 			/*
1454 			 * Only increment phys if it doesn't describe
1455 			 * a hole.
1456 			 */
1457 			phys++;
1458 		}
1459 
1460 		/*
1461 		 * If w_first_new_cpos is < UINT_MAX, we have a non-sparse
1462 		 * file that got extended.  w_first_new_cpos tells us
1463 		 * where the newly allocated clusters are so we can
1464 		 * zero them.
1465 		 */
1466 		if (desc->c_cpos >= wc->w_first_new_cpos) {
1467 			BUG_ON(phys == 0);
1468 			desc->c_needs_zero = 1;
1469 		}
1470 
1471 		desc->c_phys = phys;
1472 		if (phys == 0) {
1473 			desc->c_new = 1;
1474 			desc->c_needs_zero = 1;
1475 			desc->c_clear_unwritten = 1;
1476 			*clusters_to_alloc = *clusters_to_alloc + 1;
1477 		}
1478 
1479 		if (ext_flags & OCFS2_EXT_UNWRITTEN) {
1480 			desc->c_clear_unwritten = 1;
1481 			desc->c_needs_zero = 1;
1482 		}
1483 
1484 		ret = ocfs2_unwritten_check(inode, wc, desc);
1485 		if (ret) {
1486 			mlog_errno(ret);
1487 			goto out;
1488 		}
1489 
1490 		num_clusters--;
1491 	}
1492 
1493 	ret = 0;
1494 out:
1495 	return ret;
1496 }
1497 
1498 static int ocfs2_write_begin_inline(struct address_space *mapping,
1499 				    struct inode *inode,
1500 				    struct ocfs2_write_ctxt *wc)
1501 {
1502 	int ret;
1503 	struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1504 	struct page *page;
1505 	handle_t *handle;
1506 	struct ocfs2_dinode *di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1507 
1508 	handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1509 	if (IS_ERR(handle)) {
1510 		ret = PTR_ERR(handle);
1511 		mlog_errno(ret);
1512 		goto out;
1513 	}
1514 
1515 	page = find_or_create_page(mapping, 0, GFP_NOFS);
1516 	if (!page) {
1517 		ocfs2_commit_trans(osb, handle);
1518 		ret = -ENOMEM;
1519 		mlog_errno(ret);
1520 		goto out;
1521 	}
1522 	/*
1523 	 * If we don't set w_num_pages then this page won't get unlocked
1524 	 * and freed on cleanup of the write context.
1525 	 */
1526 	wc->w_pages[0] = wc->w_target_page = page;
1527 	wc->w_num_pages = 1;
1528 
1529 	ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), wc->w_di_bh,
1530 				      OCFS2_JOURNAL_ACCESS_WRITE);
1531 	if (ret) {
1532 		ocfs2_commit_trans(osb, handle);
1533 
1534 		mlog_errno(ret);
1535 		goto out;
1536 	}
1537 
1538 	if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL))
1539 		ocfs2_set_inode_data_inline(inode, di);
1540 
1541 	if (!PageUptodate(page)) {
1542 		ret = ocfs2_read_inline_data(inode, page, wc->w_di_bh);
1543 		if (ret) {
1544 			ocfs2_commit_trans(osb, handle);
1545 
1546 			goto out;
1547 		}
1548 	}
1549 
1550 	wc->w_handle = handle;
1551 out:
1552 	return ret;
1553 }
1554 
1555 int ocfs2_size_fits_inline_data(struct buffer_head *di_bh, u64 new_size)
1556 {
1557 	struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
1558 
1559 	if (new_size <= le16_to_cpu(di->id2.i_data.id_count))
1560 		return 1;
1561 	return 0;
1562 }
1563 
1564 static int ocfs2_try_to_write_inline_data(struct address_space *mapping,
1565 					  struct inode *inode, loff_t pos,
1566 					  unsigned len, struct page *mmap_page,
1567 					  struct ocfs2_write_ctxt *wc)
1568 {
1569 	int ret, written = 0;
1570 	loff_t end = pos + len;
1571 	struct ocfs2_inode_info *oi = OCFS2_I(inode);
1572 	struct ocfs2_dinode *di = NULL;
1573 
1574 	trace_ocfs2_try_to_write_inline_data((unsigned long long)oi->ip_blkno,
1575 					     len, (unsigned long long)pos,
1576 					     oi->ip_dyn_features);
1577 
1578 	/*
1579 	 * Handle inodes which already have inline data 1st.
1580 	 */
1581 	if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1582 		if (mmap_page == NULL &&
1583 		    ocfs2_size_fits_inline_data(wc->w_di_bh, end))
1584 			goto do_inline_write;
1585 
1586 		/*
1587 		 * The write won't fit - we have to give this inode an
1588 		 * inline extent list now.
1589 		 */
1590 		ret = ocfs2_convert_inline_data_to_extents(inode, wc->w_di_bh);
1591 		if (ret)
1592 			mlog_errno(ret);
1593 		goto out;
1594 	}
1595 
1596 	/*
1597 	 * Check whether the inode can accept inline data.
1598 	 */
1599 	if (oi->ip_clusters != 0 || i_size_read(inode) != 0)
1600 		return 0;
1601 
1602 	/*
1603 	 * Check whether the write can fit.
1604 	 */
1605 	di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1606 	if (mmap_page ||
1607 	    end > ocfs2_max_inline_data_with_xattr(inode->i_sb, di))
1608 		return 0;
1609 
1610 do_inline_write:
1611 	ret = ocfs2_write_begin_inline(mapping, inode, wc);
1612 	if (ret) {
1613 		mlog_errno(ret);
1614 		goto out;
1615 	}
1616 
1617 	/*
1618 	 * This signals to the caller that the data can be written
1619 	 * inline.
1620 	 */
1621 	written = 1;
1622 out:
1623 	return written ? written : ret;
1624 }
1625 
1626 /*
1627  * This function only does anything for file systems which can't
1628  * handle sparse files.
1629  *
1630  * What we want to do here is fill in any hole between the current end
1631  * of allocation and the end of our write. That way the rest of the
1632  * write path can treat it as an non-allocating write, which has no
1633  * special case code for sparse/nonsparse files.
1634  */
1635 static int ocfs2_expand_nonsparse_inode(struct inode *inode,
1636 					struct buffer_head *di_bh,
1637 					loff_t pos, unsigned len,
1638 					struct ocfs2_write_ctxt *wc)
1639 {
1640 	int ret;
1641 	loff_t newsize = pos + len;
1642 
1643 	BUG_ON(ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)));
1644 
1645 	if (newsize <= i_size_read(inode))
1646 		return 0;
1647 
1648 	ret = ocfs2_extend_no_holes(inode, di_bh, newsize, pos);
1649 	if (ret)
1650 		mlog_errno(ret);
1651 
1652 	/* There is no wc if this is call from direct. */
1653 	if (wc)
1654 		wc->w_first_new_cpos =
1655 			ocfs2_clusters_for_bytes(inode->i_sb, i_size_read(inode));
1656 
1657 	return ret;
1658 }
1659 
1660 static int ocfs2_zero_tail(struct inode *inode, struct buffer_head *di_bh,
1661 			   loff_t pos)
1662 {
1663 	int ret = 0;
1664 
1665 	BUG_ON(!ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)));
1666 	if (pos > i_size_read(inode))
1667 		ret = ocfs2_zero_extend(inode, di_bh, pos);
1668 
1669 	return ret;
1670 }
1671 
1672 int ocfs2_write_begin_nolock(struct address_space *mapping,
1673 			     loff_t pos, unsigned len, ocfs2_write_type_t type,
1674 			     struct page **pagep, void **fsdata,
1675 			     struct buffer_head *di_bh, struct page *mmap_page)
1676 {
1677 	int ret, cluster_of_pages, credits = OCFS2_INODE_UPDATE_CREDITS;
1678 	unsigned int clusters_to_alloc, extents_to_split, clusters_need = 0;
1679 	struct ocfs2_write_ctxt *wc;
1680 	struct inode *inode = mapping->host;
1681 	struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1682 	struct ocfs2_dinode *di;
1683 	struct ocfs2_alloc_context *data_ac = NULL;
1684 	struct ocfs2_alloc_context *meta_ac = NULL;
1685 	handle_t *handle;
1686 	struct ocfs2_extent_tree et;
1687 	int try_free = 1, ret1;
1688 
1689 try_again:
1690 	ret = ocfs2_alloc_write_ctxt(&wc, osb, pos, len, type, di_bh);
1691 	if (ret) {
1692 		mlog_errno(ret);
1693 		return ret;
1694 	}
1695 
1696 	if (ocfs2_supports_inline_data(osb)) {
1697 		ret = ocfs2_try_to_write_inline_data(mapping, inode, pos, len,
1698 						     mmap_page, wc);
1699 		if (ret == 1) {
1700 			ret = 0;
1701 			goto success;
1702 		}
1703 		if (ret < 0) {
1704 			mlog_errno(ret);
1705 			goto out;
1706 		}
1707 	}
1708 
1709 	/* Direct io change i_size late, should not zero tail here. */
1710 	if (type != OCFS2_WRITE_DIRECT) {
1711 		if (ocfs2_sparse_alloc(osb))
1712 			ret = ocfs2_zero_tail(inode, di_bh, pos);
1713 		else
1714 			ret = ocfs2_expand_nonsparse_inode(inode, di_bh, pos,
1715 							   len, wc);
1716 		if (ret) {
1717 			mlog_errno(ret);
1718 			goto out;
1719 		}
1720 	}
1721 
1722 	ret = ocfs2_check_range_for_refcount(inode, pos, len);
1723 	if (ret < 0) {
1724 		mlog_errno(ret);
1725 		goto out;
1726 	} else if (ret == 1) {
1727 		clusters_need = wc->w_clen;
1728 		ret = ocfs2_refcount_cow(inode, di_bh,
1729 					 wc->w_cpos, wc->w_clen, UINT_MAX);
1730 		if (ret) {
1731 			mlog_errno(ret);
1732 			goto out;
1733 		}
1734 	}
1735 
1736 	ret = ocfs2_populate_write_desc(inode, wc, &clusters_to_alloc,
1737 					&extents_to_split);
1738 	if (ret) {
1739 		mlog_errno(ret);
1740 		goto out;
1741 	}
1742 	clusters_need += clusters_to_alloc;
1743 
1744 	di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1745 
1746 	trace_ocfs2_write_begin_nolock(
1747 			(unsigned long long)OCFS2_I(inode)->ip_blkno,
1748 			(long long)i_size_read(inode),
1749 			le32_to_cpu(di->i_clusters),
1750 			pos, len, type, mmap_page,
1751 			clusters_to_alloc, extents_to_split);
1752 
1753 	/*
1754 	 * We set w_target_from, w_target_to here so that
1755 	 * ocfs2_write_end() knows which range in the target page to
1756 	 * write out. An allocation requires that we write the entire
1757 	 * cluster range.
1758 	 */
1759 	if (clusters_to_alloc || extents_to_split) {
1760 		/*
1761 		 * XXX: We are stretching the limits of
1762 		 * ocfs2_lock_allocators(). It greatly over-estimates
1763 		 * the work to be done.
1764 		 */
1765 		ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode),
1766 					      wc->w_di_bh);
1767 		ret = ocfs2_lock_allocators(inode, &et,
1768 					    clusters_to_alloc, extents_to_split,
1769 					    &data_ac, &meta_ac);
1770 		if (ret) {
1771 			mlog_errno(ret);
1772 			goto out;
1773 		}
1774 
1775 		if (data_ac)
1776 			data_ac->ac_resv = &OCFS2_I(inode)->ip_la_data_resv;
1777 
1778 		credits = ocfs2_calc_extend_credits(inode->i_sb,
1779 						    &di->id2.i_list);
1780 	} else if (type == OCFS2_WRITE_DIRECT)
1781 		/* direct write needs not to start trans if no extents alloc. */
1782 		goto success;
1783 
1784 	/*
1785 	 * We have to zero sparse allocated clusters, unwritten extent clusters,
1786 	 * and non-sparse clusters we just extended.  For non-sparse writes,
1787 	 * we know zeros will only be needed in the first and/or last cluster.
1788 	 */
1789 	if (wc->w_clen && (wc->w_desc[0].c_needs_zero ||
1790 			   wc->w_desc[wc->w_clen - 1].c_needs_zero))
1791 		cluster_of_pages = 1;
1792 	else
1793 		cluster_of_pages = 0;
1794 
1795 	ocfs2_set_target_boundaries(osb, wc, pos, len, cluster_of_pages);
1796 
1797 	handle = ocfs2_start_trans(osb, credits);
1798 	if (IS_ERR(handle)) {
1799 		ret = PTR_ERR(handle);
1800 		mlog_errno(ret);
1801 		goto out;
1802 	}
1803 
1804 	wc->w_handle = handle;
1805 
1806 	if (clusters_to_alloc) {
1807 		ret = dquot_alloc_space_nodirty(inode,
1808 			ocfs2_clusters_to_bytes(osb->sb, clusters_to_alloc));
1809 		if (ret)
1810 			goto out_commit;
1811 	}
1812 
1813 	ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), wc->w_di_bh,
1814 				      OCFS2_JOURNAL_ACCESS_WRITE);
1815 	if (ret) {
1816 		mlog_errno(ret);
1817 		goto out_quota;
1818 	}
1819 
1820 	/*
1821 	 * Fill our page array first. That way we've grabbed enough so
1822 	 * that we can zero and flush if we error after adding the
1823 	 * extent.
1824 	 */
1825 	ret = ocfs2_grab_pages_for_write(mapping, wc, wc->w_cpos, pos, len,
1826 					 cluster_of_pages, mmap_page);
1827 	if (ret && ret != -EAGAIN) {
1828 		mlog_errno(ret);
1829 		goto out_quota;
1830 	}
1831 
1832 	/*
1833 	 * ocfs2_grab_pages_for_write() returns -EAGAIN if it could not lock
1834 	 * the target page. In this case, we exit with no error and no target
1835 	 * page. This will trigger the caller, page_mkwrite(), to re-try
1836 	 * the operation.
1837 	 */
1838 	if (ret == -EAGAIN) {
1839 		BUG_ON(wc->w_target_page);
1840 		ret = 0;
1841 		goto out_quota;
1842 	}
1843 
1844 	ret = ocfs2_write_cluster_by_desc(mapping, data_ac, meta_ac, wc, pos,
1845 					  len);
1846 	if (ret) {
1847 		mlog_errno(ret);
1848 		goto out_quota;
1849 	}
1850 
1851 	if (data_ac)
1852 		ocfs2_free_alloc_context(data_ac);
1853 	if (meta_ac)
1854 		ocfs2_free_alloc_context(meta_ac);
1855 
1856 success:
1857 	if (pagep)
1858 		*pagep = wc->w_target_page;
1859 	*fsdata = wc;
1860 	return 0;
1861 out_quota:
1862 	if (clusters_to_alloc)
1863 		dquot_free_space(inode,
1864 			  ocfs2_clusters_to_bytes(osb->sb, clusters_to_alloc));
1865 out_commit:
1866 	ocfs2_commit_trans(osb, handle);
1867 
1868 out:
1869 	/*
1870 	 * The mmapped page won't be unlocked in ocfs2_free_write_ctxt(),
1871 	 * even in case of error here like ENOSPC and ENOMEM. So, we need
1872 	 * to unlock the target page manually to prevent deadlocks when
1873 	 * retrying again on ENOSPC, or when returning non-VM_FAULT_LOCKED
1874 	 * to VM code.
1875 	 */
1876 	if (wc->w_target_locked)
1877 		unlock_page(mmap_page);
1878 
1879 	ocfs2_free_write_ctxt(inode, wc);
1880 
1881 	if (data_ac) {
1882 		ocfs2_free_alloc_context(data_ac);
1883 		data_ac = NULL;
1884 	}
1885 	if (meta_ac) {
1886 		ocfs2_free_alloc_context(meta_ac);
1887 		meta_ac = NULL;
1888 	}
1889 
1890 	if (ret == -ENOSPC && try_free) {
1891 		/*
1892 		 * Try to free some truncate log so that we can have enough
1893 		 * clusters to allocate.
1894 		 */
1895 		try_free = 0;
1896 
1897 		ret1 = ocfs2_try_to_free_truncate_log(osb, clusters_need);
1898 		if (ret1 == 1)
1899 			goto try_again;
1900 
1901 		if (ret1 < 0)
1902 			mlog_errno(ret1);
1903 	}
1904 
1905 	return ret;
1906 }
1907 
1908 static int ocfs2_write_begin(struct file *file, struct address_space *mapping,
1909 			     loff_t pos, unsigned len, unsigned flags,
1910 			     struct page **pagep, void **fsdata)
1911 {
1912 	int ret;
1913 	struct buffer_head *di_bh = NULL;
1914 	struct inode *inode = mapping->host;
1915 
1916 	ret = ocfs2_inode_lock(inode, &di_bh, 1);
1917 	if (ret) {
1918 		mlog_errno(ret);
1919 		return ret;
1920 	}
1921 
1922 	/*
1923 	 * Take alloc sem here to prevent concurrent lookups. That way
1924 	 * the mapping, zeroing and tree manipulation within
1925 	 * ocfs2_write() will be safe against ->readpage(). This
1926 	 * should also serve to lock out allocation from a shared
1927 	 * writeable region.
1928 	 */
1929 	down_write(&OCFS2_I(inode)->ip_alloc_sem);
1930 
1931 	ret = ocfs2_write_begin_nolock(mapping, pos, len, OCFS2_WRITE_BUFFER,
1932 				       pagep, fsdata, di_bh, NULL);
1933 	if (ret) {
1934 		mlog_errno(ret);
1935 		goto out_fail;
1936 	}
1937 
1938 	brelse(di_bh);
1939 
1940 	return 0;
1941 
1942 out_fail:
1943 	up_write(&OCFS2_I(inode)->ip_alloc_sem);
1944 
1945 	brelse(di_bh);
1946 	ocfs2_inode_unlock(inode, 1);
1947 
1948 	return ret;
1949 }
1950 
1951 static void ocfs2_write_end_inline(struct inode *inode, loff_t pos,
1952 				   unsigned len, unsigned *copied,
1953 				   struct ocfs2_dinode *di,
1954 				   struct ocfs2_write_ctxt *wc)
1955 {
1956 	void *kaddr;
1957 
1958 	if (unlikely(*copied < len)) {
1959 		if (!PageUptodate(wc->w_target_page)) {
1960 			*copied = 0;
1961 			return;
1962 		}
1963 	}
1964 
1965 	kaddr = kmap_atomic(wc->w_target_page);
1966 	memcpy(di->id2.i_data.id_data + pos, kaddr + pos, *copied);
1967 	kunmap_atomic(kaddr);
1968 
1969 	trace_ocfs2_write_end_inline(
1970 	     (unsigned long long)OCFS2_I(inode)->ip_blkno,
1971 	     (unsigned long long)pos, *copied,
1972 	     le16_to_cpu(di->id2.i_data.id_count),
1973 	     le16_to_cpu(di->i_dyn_features));
1974 }
1975 
1976 int ocfs2_write_end_nolock(struct address_space *mapping,
1977 			   loff_t pos, unsigned len, unsigned copied, void *fsdata)
1978 {
1979 	int i, ret;
1980 	unsigned from, to, start = pos & (PAGE_SIZE - 1);
1981 	struct inode *inode = mapping->host;
1982 	struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1983 	struct ocfs2_write_ctxt *wc = fsdata;
1984 	struct ocfs2_dinode *di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1985 	handle_t *handle = wc->w_handle;
1986 	struct page *tmppage;
1987 
1988 	BUG_ON(!list_empty(&wc->w_unwritten_list));
1989 
1990 	if (handle) {
1991 		ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode),
1992 				wc->w_di_bh, OCFS2_JOURNAL_ACCESS_WRITE);
1993 		if (ret) {
1994 			copied = ret;
1995 			mlog_errno(ret);
1996 			goto out;
1997 		}
1998 	}
1999 
2000 	if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
2001 		ocfs2_write_end_inline(inode, pos, len, &copied, di, wc);
2002 		goto out_write_size;
2003 	}
2004 
2005 	if (unlikely(copied < len) && wc->w_target_page) {
2006 		if (!PageUptodate(wc->w_target_page))
2007 			copied = 0;
2008 
2009 		ocfs2_zero_new_buffers(wc->w_target_page, start+copied,
2010 				       start+len);
2011 	}
2012 	if (wc->w_target_page)
2013 		flush_dcache_page(wc->w_target_page);
2014 
2015 	for(i = 0; i < wc->w_num_pages; i++) {
2016 		tmppage = wc->w_pages[i];
2017 
2018 		/* This is the direct io target page. */
2019 		if (tmppage == NULL)
2020 			continue;
2021 
2022 		if (tmppage == wc->w_target_page) {
2023 			from = wc->w_target_from;
2024 			to = wc->w_target_to;
2025 
2026 			BUG_ON(from > PAGE_SIZE ||
2027 			       to > PAGE_SIZE ||
2028 			       to < from);
2029 		} else {
2030 			/*
2031 			 * Pages adjacent to the target (if any) imply
2032 			 * a hole-filling write in which case we want
2033 			 * to flush their entire range.
2034 			 */
2035 			from = 0;
2036 			to = PAGE_SIZE;
2037 		}
2038 
2039 		if (page_has_buffers(tmppage)) {
2040 			if (handle && ocfs2_should_order_data(inode))
2041 				ocfs2_jbd2_file_inode(handle, inode);
2042 			block_commit_write(tmppage, from, to);
2043 		}
2044 	}
2045 
2046 out_write_size:
2047 	/* Direct io do not update i_size here. */
2048 	if (wc->w_type != OCFS2_WRITE_DIRECT) {
2049 		pos += copied;
2050 		if (pos > i_size_read(inode)) {
2051 			i_size_write(inode, pos);
2052 			mark_inode_dirty(inode);
2053 		}
2054 		inode->i_blocks = ocfs2_inode_sector_count(inode);
2055 		di->i_size = cpu_to_le64((u64)i_size_read(inode));
2056 		inode->i_mtime = inode->i_ctime = current_time(inode);
2057 		di->i_mtime = di->i_ctime = cpu_to_le64(inode->i_mtime.tv_sec);
2058 		di->i_mtime_nsec = di->i_ctime_nsec = cpu_to_le32(inode->i_mtime.tv_nsec);
2059 		ocfs2_update_inode_fsync_trans(handle, inode, 1);
2060 	}
2061 	if (handle)
2062 		ocfs2_journal_dirty(handle, wc->w_di_bh);
2063 
2064 out:
2065 	/* unlock pages before dealloc since it needs acquiring j_trans_barrier
2066 	 * lock, or it will cause a deadlock since journal commit threads holds
2067 	 * this lock and will ask for the page lock when flushing the data.
2068 	 * put it here to preserve the unlock order.
2069 	 */
2070 	ocfs2_unlock_pages(wc);
2071 
2072 	if (handle)
2073 		ocfs2_commit_trans(osb, handle);
2074 
2075 	ocfs2_run_deallocs(osb, &wc->w_dealloc);
2076 
2077 	brelse(wc->w_di_bh);
2078 	kfree(wc);
2079 
2080 	return copied;
2081 }
2082 
2083 static int ocfs2_write_end(struct file *file, struct address_space *mapping,
2084 			   loff_t pos, unsigned len, unsigned copied,
2085 			   struct page *page, void *fsdata)
2086 {
2087 	int ret;
2088 	struct inode *inode = mapping->host;
2089 
2090 	ret = ocfs2_write_end_nolock(mapping, pos, len, copied, fsdata);
2091 
2092 	up_write(&OCFS2_I(inode)->ip_alloc_sem);
2093 	ocfs2_inode_unlock(inode, 1);
2094 
2095 	return ret;
2096 }
2097 
2098 struct ocfs2_dio_write_ctxt {
2099 	struct list_head	dw_zero_list;
2100 	unsigned		dw_zero_count;
2101 	int			dw_orphaned;
2102 	pid_t			dw_writer_pid;
2103 };
2104 
2105 static struct ocfs2_dio_write_ctxt *
2106 ocfs2_dio_alloc_write_ctx(struct buffer_head *bh, int *alloc)
2107 {
2108 	struct ocfs2_dio_write_ctxt *dwc = NULL;
2109 
2110 	if (bh->b_private)
2111 		return bh->b_private;
2112 
2113 	dwc = kmalloc(sizeof(struct ocfs2_dio_write_ctxt), GFP_NOFS);
2114 	if (dwc == NULL)
2115 		return NULL;
2116 	INIT_LIST_HEAD(&dwc->dw_zero_list);
2117 	dwc->dw_zero_count = 0;
2118 	dwc->dw_orphaned = 0;
2119 	dwc->dw_writer_pid = task_pid_nr(current);
2120 	bh->b_private = dwc;
2121 	*alloc = 1;
2122 
2123 	return dwc;
2124 }
2125 
2126 static void ocfs2_dio_free_write_ctx(struct inode *inode,
2127 				     struct ocfs2_dio_write_ctxt *dwc)
2128 {
2129 	ocfs2_free_unwritten_list(inode, &dwc->dw_zero_list);
2130 	kfree(dwc);
2131 }
2132 
2133 /*
2134  * TODO: Make this into a generic get_blocks function.
2135  *
2136  * From do_direct_io in direct-io.c:
2137  *  "So what we do is to permit the ->get_blocks function to populate
2138  *   bh.b_size with the size of IO which is permitted at this offset and
2139  *   this i_blkbits."
2140  *
2141  * This function is called directly from get_more_blocks in direct-io.c.
2142  *
2143  * called like this: dio->get_blocks(dio->inode, fs_startblk,
2144  * 					fs_count, map_bh, dio->rw == WRITE);
2145  */
2146 static int ocfs2_dio_wr_get_block(struct inode *inode, sector_t iblock,
2147 			       struct buffer_head *bh_result, int create)
2148 {
2149 	struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
2150 	struct ocfs2_inode_info *oi = OCFS2_I(inode);
2151 	struct ocfs2_write_ctxt *wc;
2152 	struct ocfs2_write_cluster_desc *desc = NULL;
2153 	struct ocfs2_dio_write_ctxt *dwc = NULL;
2154 	struct buffer_head *di_bh = NULL;
2155 	u64 p_blkno;
2156 	loff_t pos = iblock << inode->i_sb->s_blocksize_bits;
2157 	unsigned len, total_len = bh_result->b_size;
2158 	int ret = 0, first_get_block = 0;
2159 
2160 	len = osb->s_clustersize - (pos & (osb->s_clustersize - 1));
2161 	len = min(total_len, len);
2162 
2163 	mlog(0, "get block of %lu at %llu:%u req %u\n",
2164 			inode->i_ino, pos, len, total_len);
2165 
2166 	/*
2167 	 * Because we need to change file size in ocfs2_dio_end_io_write(), or
2168 	 * we may need to add it to orphan dir. So can not fall to fast path
2169 	 * while file size will be changed.
2170 	 */
2171 	if (pos + total_len <= i_size_read(inode)) {
2172 
2173 		/* This is the fast path for re-write. */
2174 		ret = ocfs2_lock_get_block(inode, iblock, bh_result, create);
2175 		if (buffer_mapped(bh_result) &&
2176 		    !buffer_new(bh_result) &&
2177 		    ret == 0)
2178 			goto out;
2179 
2180 		/* Clear state set by ocfs2_get_block. */
2181 		bh_result->b_state = 0;
2182 	}
2183 
2184 	dwc = ocfs2_dio_alloc_write_ctx(bh_result, &first_get_block);
2185 	if (unlikely(dwc == NULL)) {
2186 		ret = -ENOMEM;
2187 		mlog_errno(ret);
2188 		goto out;
2189 	}
2190 
2191 	if (ocfs2_clusters_for_bytes(inode->i_sb, pos + total_len) >
2192 	    ocfs2_clusters_for_bytes(inode->i_sb, i_size_read(inode)) &&
2193 	    !dwc->dw_orphaned) {
2194 		/*
2195 		 * when we are going to alloc extents beyond file size, add the
2196 		 * inode to orphan dir, so we can recall those spaces when
2197 		 * system crashed during write.
2198 		 */
2199 		ret = ocfs2_add_inode_to_orphan(osb, inode);
2200 		if (ret < 0) {
2201 			mlog_errno(ret);
2202 			goto out;
2203 		}
2204 		dwc->dw_orphaned = 1;
2205 	}
2206 
2207 	ret = ocfs2_inode_lock(inode, &di_bh, 1);
2208 	if (ret) {
2209 		mlog_errno(ret);
2210 		goto out;
2211 	}
2212 
2213 	down_write(&oi->ip_alloc_sem);
2214 
2215 	if (first_get_block) {
2216 		if (ocfs2_sparse_alloc(osb))
2217 			ret = ocfs2_zero_tail(inode, di_bh, pos);
2218 		else
2219 			ret = ocfs2_expand_nonsparse_inode(inode, di_bh, pos,
2220 							   total_len, NULL);
2221 		if (ret < 0) {
2222 			mlog_errno(ret);
2223 			goto unlock;
2224 		}
2225 	}
2226 
2227 	ret = ocfs2_write_begin_nolock(inode->i_mapping, pos, len,
2228 				       OCFS2_WRITE_DIRECT, NULL,
2229 				       (void **)&wc, di_bh, NULL);
2230 	if (ret) {
2231 		mlog_errno(ret);
2232 		goto unlock;
2233 	}
2234 
2235 	desc = &wc->w_desc[0];
2236 
2237 	p_blkno = ocfs2_clusters_to_blocks(inode->i_sb, desc->c_phys);
2238 	BUG_ON(p_blkno == 0);
2239 	p_blkno += iblock & (u64)(ocfs2_clusters_to_blocks(inode->i_sb, 1) - 1);
2240 
2241 	map_bh(bh_result, inode->i_sb, p_blkno);
2242 	bh_result->b_size = len;
2243 	if (desc->c_needs_zero)
2244 		set_buffer_new(bh_result);
2245 
2246 	/* May sleep in end_io. It should not happen in a irq context. So defer
2247 	 * it to dio work queue. */
2248 	set_buffer_defer_completion(bh_result);
2249 
2250 	if (!list_empty(&wc->w_unwritten_list)) {
2251 		struct ocfs2_unwritten_extent *ue = NULL;
2252 
2253 		ue = list_first_entry(&wc->w_unwritten_list,
2254 				      struct ocfs2_unwritten_extent,
2255 				      ue_node);
2256 		BUG_ON(ue->ue_cpos != desc->c_cpos);
2257 		/* The physical address may be 0, fill it. */
2258 		ue->ue_phys = desc->c_phys;
2259 
2260 		list_splice_tail_init(&wc->w_unwritten_list, &dwc->dw_zero_list);
2261 		dwc->dw_zero_count += wc->w_unwritten_count;
2262 	}
2263 
2264 	ret = ocfs2_write_end_nolock(inode->i_mapping, pos, len, len, wc);
2265 	BUG_ON(ret != len);
2266 	ret = 0;
2267 unlock:
2268 	up_write(&oi->ip_alloc_sem);
2269 	ocfs2_inode_unlock(inode, 1);
2270 	brelse(di_bh);
2271 out:
2272 	if (ret < 0)
2273 		ret = -EIO;
2274 	return ret;
2275 }
2276 
2277 static int ocfs2_dio_end_io_write(struct inode *inode,
2278 				  struct ocfs2_dio_write_ctxt *dwc,
2279 				  loff_t offset,
2280 				  ssize_t bytes)
2281 {
2282 	struct ocfs2_cached_dealloc_ctxt dealloc;
2283 	struct ocfs2_extent_tree et;
2284 	struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
2285 	struct ocfs2_inode_info *oi = OCFS2_I(inode);
2286 	struct ocfs2_unwritten_extent *ue = NULL;
2287 	struct buffer_head *di_bh = NULL;
2288 	struct ocfs2_dinode *di;
2289 	struct ocfs2_alloc_context *data_ac = NULL;
2290 	struct ocfs2_alloc_context *meta_ac = NULL;
2291 	handle_t *handle = NULL;
2292 	loff_t end = offset + bytes;
2293 	int ret = 0, credits = 0, locked = 0;
2294 
2295 	ocfs2_init_dealloc_ctxt(&dealloc);
2296 
2297 	/* We do clear unwritten, delete orphan, change i_size here. If neither
2298 	 * of these happen, we can skip all this. */
2299 	if (list_empty(&dwc->dw_zero_list) &&
2300 	    end <= i_size_read(inode) &&
2301 	    !dwc->dw_orphaned)
2302 		goto out;
2303 
2304 	/* ocfs2_file_write_iter will get i_mutex, so we need not lock if we
2305 	 * are in that context. */
2306 	if (dwc->dw_writer_pid != task_pid_nr(current)) {
2307 		inode_lock(inode);
2308 		locked = 1;
2309 	}
2310 
2311 	ret = ocfs2_inode_lock(inode, &di_bh, 1);
2312 	if (ret < 0) {
2313 		mlog_errno(ret);
2314 		goto out;
2315 	}
2316 
2317 	down_write(&oi->ip_alloc_sem);
2318 
2319 	/* Delete orphan before acquire i_mutex. */
2320 	if (dwc->dw_orphaned) {
2321 		BUG_ON(dwc->dw_writer_pid != task_pid_nr(current));
2322 
2323 		end = end > i_size_read(inode) ? end : 0;
2324 
2325 		ret = ocfs2_del_inode_from_orphan(osb, inode, di_bh,
2326 				!!end, end);
2327 		if (ret < 0)
2328 			mlog_errno(ret);
2329 	}
2330 
2331 	di = (struct ocfs2_dinode *)di_bh->b_data;
2332 
2333 	ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), di_bh);
2334 
2335 	/* Attach dealloc with extent tree in case that we may reuse extents
2336 	 * which are already unlinked from current extent tree due to extent
2337 	 * rotation and merging.
2338 	 */
2339 	et.et_dealloc = &dealloc;
2340 
2341 	ret = ocfs2_lock_allocators(inode, &et, 0, dwc->dw_zero_count*2,
2342 				    &data_ac, &meta_ac);
2343 	if (ret) {
2344 		mlog_errno(ret);
2345 		goto unlock;
2346 	}
2347 
2348 	credits = ocfs2_calc_extend_credits(inode->i_sb, &di->id2.i_list);
2349 
2350 	handle = ocfs2_start_trans(osb, credits);
2351 	if (IS_ERR(handle)) {
2352 		ret = PTR_ERR(handle);
2353 		mlog_errno(ret);
2354 		goto unlock;
2355 	}
2356 	ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
2357 				      OCFS2_JOURNAL_ACCESS_WRITE);
2358 	if (ret) {
2359 		mlog_errno(ret);
2360 		goto commit;
2361 	}
2362 
2363 	list_for_each_entry(ue, &dwc->dw_zero_list, ue_node) {
2364 		ret = ocfs2_mark_extent_written(inode, &et, handle,
2365 						ue->ue_cpos, 1,
2366 						ue->ue_phys,
2367 						meta_ac, &dealloc);
2368 		if (ret < 0) {
2369 			mlog_errno(ret);
2370 			break;
2371 		}
2372 	}
2373 
2374 	if (end > i_size_read(inode)) {
2375 		ret = ocfs2_set_inode_size(handle, inode, di_bh, end);
2376 		if (ret < 0)
2377 			mlog_errno(ret);
2378 	}
2379 commit:
2380 	ocfs2_commit_trans(osb, handle);
2381 unlock:
2382 	up_write(&oi->ip_alloc_sem);
2383 	ocfs2_inode_unlock(inode, 1);
2384 	brelse(di_bh);
2385 out:
2386 	if (data_ac)
2387 		ocfs2_free_alloc_context(data_ac);
2388 	if (meta_ac)
2389 		ocfs2_free_alloc_context(meta_ac);
2390 	ocfs2_run_deallocs(osb, &dealloc);
2391 	if (locked)
2392 		inode_unlock(inode);
2393 	ocfs2_dio_free_write_ctx(inode, dwc);
2394 
2395 	return ret;
2396 }
2397 
2398 /*
2399  * ocfs2_dio_end_io is called by the dio core when a dio is finished.  We're
2400  * particularly interested in the aio/dio case.  We use the rw_lock DLM lock
2401  * to protect io on one node from truncation on another.
2402  */
2403 static int ocfs2_dio_end_io(struct kiocb *iocb,
2404 			    loff_t offset,
2405 			    ssize_t bytes,
2406 			    void *private)
2407 {
2408 	struct inode *inode = file_inode(iocb->ki_filp);
2409 	int level;
2410 	int ret = 0;
2411 
2412 	/* this io's submitter should not have unlocked this before we could */
2413 	BUG_ON(!ocfs2_iocb_is_rw_locked(iocb));
2414 
2415 	if (bytes <= 0)
2416 		mlog_ratelimited(ML_ERROR, "Direct IO failed, bytes = %lld",
2417 				 (long long)bytes);
2418 	if (private) {
2419 		if (bytes > 0)
2420 			ret = ocfs2_dio_end_io_write(inode, private, offset,
2421 						     bytes);
2422 		else
2423 			ocfs2_dio_free_write_ctx(inode, private);
2424 	}
2425 
2426 	ocfs2_iocb_clear_rw_locked(iocb);
2427 
2428 	level = ocfs2_iocb_rw_locked_level(iocb);
2429 	ocfs2_rw_unlock(inode, level);
2430 	return ret;
2431 }
2432 
2433 static ssize_t ocfs2_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
2434 {
2435 	struct file *file = iocb->ki_filp;
2436 	struct inode *inode = file->f_mapping->host;
2437 	struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
2438 	get_block_t *get_block;
2439 
2440 	/*
2441 	 * Fallback to buffered I/O if we see an inode without
2442 	 * extents.
2443 	 */
2444 	if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
2445 		return 0;
2446 
2447 	/* Fallback to buffered I/O if we do not support append dio. */
2448 	if (iocb->ki_pos + iter->count > i_size_read(inode) &&
2449 	    !ocfs2_supports_append_dio(osb))
2450 		return 0;
2451 
2452 	if (iov_iter_rw(iter) == READ)
2453 		get_block = ocfs2_lock_get_block;
2454 	else
2455 		get_block = ocfs2_dio_wr_get_block;
2456 
2457 	return __blockdev_direct_IO(iocb, inode, inode->i_sb->s_bdev,
2458 				    iter, get_block,
2459 				    ocfs2_dio_end_io, NULL, 0);
2460 }
2461 
2462 const struct address_space_operations ocfs2_aops = {
2463 	.readpage		= ocfs2_readpage,
2464 	.readpages		= ocfs2_readpages,
2465 	.writepage		= ocfs2_writepage,
2466 	.write_begin		= ocfs2_write_begin,
2467 	.write_end		= ocfs2_write_end,
2468 	.bmap			= ocfs2_bmap,
2469 	.direct_IO		= ocfs2_direct_IO,
2470 	.invalidatepage		= block_invalidatepage,
2471 	.releasepage		= ocfs2_releasepage,
2472 	.migratepage		= buffer_migrate_page,
2473 	.is_partially_uptodate	= block_is_partially_uptodate,
2474 	.error_remove_page	= generic_error_remove_page,
2475 };
2476