1 /* -*- mode: c; c-basic-offset: 8; -*-
2 * vim: noexpandtab sw=8 ts=8 sts=0:
4 * Copyright (C) 2002, 2004 Oracle. All rights reserved.
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.
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.
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.
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>
33 #include <cluster/masklog.h>
40 #include "extent_map.h"
47 #include "refcounttree.h"
48 #include "ocfs2_trace.h"
50 #include "buffer_head_io.h"
55 static int ocfs2_symlink_get_block(struct inode *inode, sector_t iblock,
56 struct buffer_head *bh_result, int create)
60 struct ocfs2_dinode *fe = NULL;
61 struct buffer_head *bh = NULL;
62 struct buffer_head *buffer_cache_bh = NULL;
63 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
66 trace_ocfs2_symlink_get_block(
67 (unsigned long long)OCFS2_I(inode)->ip_blkno,
68 (unsigned long long)iblock, bh_result, create);
70 BUG_ON(ocfs2_inode_is_fast_symlink(inode));
72 if ((iblock << inode->i_sb->s_blocksize_bits) > PATH_MAX + 1) {
73 mlog(ML_ERROR, "block offset > PATH_MAX: %llu",
74 (unsigned long long)iblock);
78 status = ocfs2_read_inode_block(inode, &bh);
83 fe = (struct ocfs2_dinode *) bh->b_data;
85 if ((u64)iblock >= ocfs2_clusters_to_blocks(inode->i_sb,
86 le32_to_cpu(fe->i_clusters))) {
88 mlog(ML_ERROR, "block offset is outside the allocated size: "
89 "%llu\n", (unsigned long long)iblock);
93 /* We don't use the page cache to create symlink data, so if
94 * need be, copy it over from the buffer cache. */
95 if (!buffer_uptodate(bh_result) && ocfs2_inode_is_new(inode)) {
96 u64 blkno = le64_to_cpu(fe->id2.i_list.l_recs[0].e_blkno) +
98 buffer_cache_bh = sb_getblk(osb->sb, blkno);
99 if (!buffer_cache_bh) {
101 mlog(ML_ERROR, "couldn't getblock for symlink!\n");
105 /* we haven't locked out transactions, so a commit
106 * could've happened. Since we've got a reference on
107 * the bh, even if it commits while we're doing the
108 * copy, the data is still good. */
109 if (buffer_jbd(buffer_cache_bh)
110 && ocfs2_inode_is_new(inode)) {
111 kaddr = kmap_atomic(bh_result->b_page);
113 mlog(ML_ERROR, "couldn't kmap!\n");
116 memcpy(kaddr + (bh_result->b_size * iblock),
117 buffer_cache_bh->b_data,
119 kunmap_atomic(kaddr);
120 set_buffer_uptodate(bh_result);
122 brelse(buffer_cache_bh);
125 map_bh(bh_result, inode->i_sb,
126 le64_to_cpu(fe->id2.i_list.l_recs[0].e_blkno) + iblock);
136 int ocfs2_get_block(struct inode *inode, sector_t iblock,
137 struct buffer_head *bh_result, int create)
140 unsigned int ext_flags;
141 u64 max_blocks = bh_result->b_size >> inode->i_blkbits;
142 u64 p_blkno, count, past_eof;
143 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
145 trace_ocfs2_get_block((unsigned long long)OCFS2_I(inode)->ip_blkno,
146 (unsigned long long)iblock, bh_result, create);
148 if (OCFS2_I(inode)->ip_flags & OCFS2_INODE_SYSTEM_FILE)
149 mlog(ML_NOTICE, "get_block on system inode 0x%p (%lu)\n",
150 inode, inode->i_ino);
152 if (S_ISLNK(inode->i_mode)) {
153 /* this always does I/O for some reason. */
154 err = ocfs2_symlink_get_block(inode, iblock, bh_result, create);
158 err = ocfs2_extent_map_get_blocks(inode, iblock, &p_blkno, &count,
161 mlog(ML_ERROR, "Error %d from get_blocks(0x%p, %llu, 1, "
162 "%llu, NULL)\n", err, inode, (unsigned long long)iblock,
163 (unsigned long long)p_blkno);
167 if (max_blocks < count)
171 * ocfs2 never allocates in this function - the only time we
172 * need to use BH_New is when we're extending i_size on a file
173 * system which doesn't support holes, in which case BH_New
174 * allows __block_write_begin() to zero.
176 * If we see this on a sparse file system, then a truncate has
177 * raced us and removed the cluster. In this case, we clear
178 * the buffers dirty and uptodate bits and let the buffer code
179 * ignore it as a hole.
181 if (create && p_blkno == 0 && ocfs2_sparse_alloc(osb)) {
182 clear_buffer_dirty(bh_result);
183 clear_buffer_uptodate(bh_result);
187 /* Treat the unwritten extent as a hole for zeroing purposes. */
188 if (p_blkno && !(ext_flags & OCFS2_EXT_UNWRITTEN))
189 map_bh(bh_result, inode->i_sb, p_blkno);
191 bh_result->b_size = count << inode->i_blkbits;
193 if (!ocfs2_sparse_alloc(osb)) {
197 "iblock = %llu p_blkno = %llu blkno=(%llu)\n",
198 (unsigned long long)iblock,
199 (unsigned long long)p_blkno,
200 (unsigned long long)OCFS2_I(inode)->ip_blkno);
201 mlog(ML_ERROR, "Size %llu, clusters %u\n", (unsigned long long)i_size_read(inode), OCFS2_I(inode)->ip_clusters);
207 past_eof = ocfs2_blocks_for_bytes(inode->i_sb, i_size_read(inode));
209 trace_ocfs2_get_block_end((unsigned long long)OCFS2_I(inode)->ip_blkno,
210 (unsigned long long)past_eof);
211 if (create && (iblock >= past_eof))
212 set_buffer_new(bh_result);
221 int ocfs2_read_inline_data(struct inode *inode, struct page *page,
222 struct buffer_head *di_bh)
226 struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
228 if (!(le16_to_cpu(di->i_dyn_features) & OCFS2_INLINE_DATA_FL)) {
229 ocfs2_error(inode->i_sb, "Inode %llu lost inline data flag",
230 (unsigned long long)OCFS2_I(inode)->ip_blkno);
234 size = i_size_read(inode);
236 if (size > PAGE_CACHE_SIZE ||
237 size > ocfs2_max_inline_data_with_xattr(inode->i_sb, di)) {
238 ocfs2_error(inode->i_sb,
239 "Inode %llu has with inline data has bad size: %Lu",
240 (unsigned long long)OCFS2_I(inode)->ip_blkno,
241 (unsigned long long)size);
245 kaddr = kmap_atomic(page);
247 memcpy(kaddr, di->id2.i_data.id_data, size);
248 /* Clear the remaining part of the page */
249 memset(kaddr + size, 0, PAGE_CACHE_SIZE - size);
250 flush_dcache_page(page);
251 kunmap_atomic(kaddr);
253 SetPageUptodate(page);
258 static int ocfs2_readpage_inline(struct inode *inode, struct page *page)
261 struct buffer_head *di_bh = NULL;
263 BUG_ON(!PageLocked(page));
264 BUG_ON(!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL));
266 ret = ocfs2_read_inode_block(inode, &di_bh);
272 ret = ocfs2_read_inline_data(inode, page, di_bh);
280 static int ocfs2_readpage(struct file *file, struct page *page)
282 struct inode *inode = page->mapping->host;
283 struct ocfs2_inode_info *oi = OCFS2_I(inode);
284 loff_t start = (loff_t)page->index << PAGE_CACHE_SHIFT;
287 trace_ocfs2_readpage((unsigned long long)oi->ip_blkno,
288 (page ? page->index : 0));
290 ret = ocfs2_inode_lock_with_page(inode, NULL, 0, page);
292 if (ret == AOP_TRUNCATED_PAGE)
298 if (down_read_trylock(&oi->ip_alloc_sem) == 0) {
300 * Unlock the page and cycle ip_alloc_sem so that we don't
301 * busyloop waiting for ip_alloc_sem to unlock
303 ret = AOP_TRUNCATED_PAGE;
306 down_read(&oi->ip_alloc_sem);
307 up_read(&oi->ip_alloc_sem);
308 goto out_inode_unlock;
312 * i_size might have just been updated as we grabed the meta lock. We
313 * might now be discovering a truncate that hit on another node.
314 * block_read_full_page->get_block freaks out if it is asked to read
315 * beyond the end of a file, so we check here. Callers
316 * (generic_file_read, vm_ops->fault) are clever enough to check i_size
317 * and notice that the page they just read isn't needed.
319 * XXX sys_readahead() seems to get that wrong?
321 if (start >= i_size_read(inode)) {
322 zero_user(page, 0, PAGE_SIZE);
323 SetPageUptodate(page);
328 if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL)
329 ret = ocfs2_readpage_inline(inode, page);
331 ret = block_read_full_page(page, ocfs2_get_block);
335 up_read(&OCFS2_I(inode)->ip_alloc_sem);
337 ocfs2_inode_unlock(inode, 0);
345 * This is used only for read-ahead. Failures or difficult to handle
346 * situations are safe to ignore.
348 * Right now, we don't bother with BH_Boundary - in-inode extent lists
349 * are quite large (243 extents on 4k blocks), so most inodes don't
350 * grow out to a tree. If need be, detecting boundary extents could
351 * trivially be added in a future version of ocfs2_get_block().
353 static int ocfs2_readpages(struct file *filp, struct address_space *mapping,
354 struct list_head *pages, unsigned nr_pages)
357 struct inode *inode = mapping->host;
358 struct ocfs2_inode_info *oi = OCFS2_I(inode);
363 * Use the nonblocking flag for the dlm code to avoid page
364 * lock inversion, but don't bother with retrying.
366 ret = ocfs2_inode_lock_full(inode, NULL, 0, OCFS2_LOCK_NONBLOCK);
370 if (down_read_trylock(&oi->ip_alloc_sem) == 0) {
371 ocfs2_inode_unlock(inode, 0);
376 * Don't bother with inline-data. There isn't anything
377 * to read-ahead in that case anyway...
379 if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL)
383 * Check whether a remote node truncated this file - we just
384 * drop out in that case as it's not worth handling here.
386 last = list_entry(pages->prev, struct page, lru);
387 start = (loff_t)last->index << PAGE_CACHE_SHIFT;
388 if (start >= i_size_read(inode))
391 err = mpage_readpages(mapping, pages, nr_pages, ocfs2_get_block);
394 up_read(&oi->ip_alloc_sem);
395 ocfs2_inode_unlock(inode, 0);
400 /* Note: Because we don't support holes, our allocation has
401 * already happened (allocation writes zeros to the file data)
402 * so we don't have to worry about ordered writes in
405 * ->writepage is called during the process of invalidating the page cache
406 * during blocked lock processing. It can't block on any cluster locks
407 * to during block mapping. It's relying on the fact that the block
408 * mapping can't have disappeared under the dirty pages that it is
409 * being asked to write back.
411 static int ocfs2_writepage(struct page *page, struct writeback_control *wbc)
413 trace_ocfs2_writepage(
414 (unsigned long long)OCFS2_I(page->mapping->host)->ip_blkno,
417 return block_write_full_page(page, ocfs2_get_block, wbc);
420 /* Taken from ext3. We don't necessarily need the full blown
421 * functionality yet, but IMHO it's better to cut and paste the whole
422 * thing so we can avoid introducing our own bugs (and easily pick up
423 * their fixes when they happen) --Mark */
424 int walk_page_buffers( handle_t *handle,
425 struct buffer_head *head,
429 int (*fn)( handle_t *handle,
430 struct buffer_head *bh))
432 struct buffer_head *bh;
433 unsigned block_start, block_end;
434 unsigned blocksize = head->b_size;
436 struct buffer_head *next;
438 for ( bh = head, block_start = 0;
439 ret == 0 && (bh != head || !block_start);
440 block_start = block_end, bh = next)
442 next = bh->b_this_page;
443 block_end = block_start + blocksize;
444 if (block_end <= from || block_start >= to) {
445 if (partial && !buffer_uptodate(bh))
449 err = (*fn)(handle, bh);
456 static sector_t ocfs2_bmap(struct address_space *mapping, sector_t block)
461 struct inode *inode = mapping->host;
463 trace_ocfs2_bmap((unsigned long long)OCFS2_I(inode)->ip_blkno,
464 (unsigned long long)block);
466 /* We don't need to lock journal system files, since they aren't
467 * accessed concurrently from multiple nodes.
469 if (!INODE_JOURNAL(inode)) {
470 err = ocfs2_inode_lock(inode, NULL, 0);
476 down_read(&OCFS2_I(inode)->ip_alloc_sem);
479 if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL))
480 err = ocfs2_extent_map_get_blocks(inode, block, &p_blkno, NULL,
483 if (!INODE_JOURNAL(inode)) {
484 up_read(&OCFS2_I(inode)->ip_alloc_sem);
485 ocfs2_inode_unlock(inode, 0);
489 mlog(ML_ERROR, "get_blocks() failed, block = %llu\n",
490 (unsigned long long)block);
496 status = err ? 0 : p_blkno;
502 * TODO: Make this into a generic get_blocks function.
504 * From do_direct_io in direct-io.c:
505 * "So what we do is to permit the ->get_blocks function to populate
506 * bh.b_size with the size of IO which is permitted at this offset and
509 * This function is called directly from get_more_blocks in direct-io.c.
511 * called like this: dio->get_blocks(dio->inode, fs_startblk,
512 * fs_count, map_bh, dio->rw == WRITE);
514 static int ocfs2_direct_IO_get_blocks(struct inode *inode, sector_t iblock,
515 struct buffer_head *bh_result, int create)
519 int alloc_locked = 0;
520 u64 p_blkno, inode_blocks, contig_blocks;
521 unsigned int ext_flags;
522 unsigned char blocksize_bits = inode->i_sb->s_blocksize_bits;
523 unsigned long max_blocks = bh_result->b_size >> inode->i_blkbits;
524 unsigned long len = bh_result->b_size;
525 unsigned int clusters_to_alloc = 0;
527 cpos = ocfs2_blocks_to_clusters(inode->i_sb, iblock);
529 /* This function won't even be called if the request isn't all
530 * nicely aligned and of the right size, so there's no need
531 * for us to check any of that. */
533 inode_blocks = ocfs2_blocks_for_bytes(inode->i_sb, i_size_read(inode));
535 /* This figures out the size of the next contiguous block, and
536 * our logical offset */
537 ret = ocfs2_extent_map_get_blocks(inode, iblock, &p_blkno,
538 &contig_blocks, &ext_flags);
540 mlog(ML_ERROR, "get_blocks() failed iblock=%llu\n",
541 (unsigned long long)iblock);
546 /* We should already CoW the refcounted extent in case of create. */
547 BUG_ON(create && (ext_flags & OCFS2_EXT_REFCOUNTED));
549 /* allocate blocks if no p_blkno is found, and create == 1 */
550 if (!p_blkno && create) {
551 ret = ocfs2_inode_lock(inode, NULL, 1);
559 /* fill hole, allocate blocks can't be larger than the size
561 clusters_to_alloc = ocfs2_clusters_for_bytes(inode->i_sb, len);
562 if (clusters_to_alloc > contig_blocks)
563 clusters_to_alloc = contig_blocks;
565 /* allocate extent and insert them into the extent tree */
566 ret = ocfs2_extend_allocation(inode, cpos,
567 clusters_to_alloc, 0);
573 ret = ocfs2_extent_map_get_blocks(inode, iblock, &p_blkno,
574 &contig_blocks, &ext_flags);
576 mlog(ML_ERROR, "get_blocks() failed iblock=%llu\n",
577 (unsigned long long)iblock);
584 * get_more_blocks() expects us to describe a hole by clearing
585 * the mapped bit on bh_result().
587 * Consider an unwritten extent as a hole.
589 if (p_blkno && !(ext_flags & OCFS2_EXT_UNWRITTEN))
590 map_bh(bh_result, inode->i_sb, p_blkno);
592 clear_buffer_mapped(bh_result);
594 /* make sure we don't map more than max_blocks blocks here as
595 that's all the kernel will handle at this point. */
596 if (max_blocks < contig_blocks)
597 contig_blocks = max_blocks;
598 bh_result->b_size = contig_blocks << blocksize_bits;
601 ocfs2_inode_unlock(inode, 1);
606 * ocfs2_dio_end_io is called by the dio core when a dio is finished. We're
607 * particularly interested in the aio/dio case. We use the rw_lock DLM lock
608 * to protect io on one node from truncation on another.
610 static void ocfs2_dio_end_io(struct kiocb *iocb,
615 struct inode *inode = file_inode(iocb->ki_filp);
618 /* this io's submitter should not have unlocked this before we could */
619 BUG_ON(!ocfs2_iocb_is_rw_locked(iocb));
621 if (ocfs2_iocb_is_sem_locked(iocb))
622 ocfs2_iocb_clear_sem_locked(iocb);
624 if (ocfs2_iocb_is_unaligned_aio(iocb)) {
625 ocfs2_iocb_clear_unaligned_aio(iocb);
627 mutex_unlock(&OCFS2_I(inode)->ip_unaligned_aio);
630 ocfs2_iocb_clear_rw_locked(iocb);
632 level = ocfs2_iocb_rw_locked_level(iocb);
633 ocfs2_rw_unlock(inode, level);
636 static int ocfs2_releasepage(struct page *page, gfp_t wait)
638 if (!page_has_buffers(page))
640 return try_to_free_buffers(page);
643 static int ocfs2_is_overwrite(struct ocfs2_super *osb,
644 struct inode *inode, loff_t offset)
649 unsigned int num_clusters = 0;
650 unsigned int ext_flags = 0;
652 v_cpos = ocfs2_bytes_to_clusters(osb->sb, offset);
653 ret = ocfs2_get_clusters(inode, v_cpos, &p_cpos,
654 &num_clusters, &ext_flags);
660 if (p_cpos && !(ext_flags & OCFS2_EXT_UNWRITTEN))
666 static int ocfs2_direct_IO_zero_extend(struct ocfs2_super *osb,
667 struct inode *inode, loff_t offset,
668 u64 zero_len, int cluster_align)
671 u32 v_cpos = ocfs2_bytes_to_clusters(osb->sb, i_size_read(inode));
672 unsigned int num_clusters = 0;
673 unsigned int ext_flags = 0;
676 if (offset <= i_size_read(inode) || cluster_align)
679 ret = ocfs2_get_clusters(inode, v_cpos, &p_cpos, &num_clusters,
686 if (p_cpos && !(ext_flags & OCFS2_EXT_UNWRITTEN)) {
687 u64 s = i_size_read(inode);
688 sector_t sector = (p_cpos << (osb->s_clustersize_bits - 9)) +
689 (do_div(s, osb->s_clustersize) >> 9);
691 ret = blkdev_issue_zeroout(osb->sb->s_bdev, sector,
692 zero_len >> 9, GFP_NOFS, false);
700 static int ocfs2_direct_IO_extend_no_holes(struct ocfs2_super *osb,
701 struct inode *inode, loff_t offset)
703 u64 zero_start, zero_len, total_zero_len;
704 u32 p_cpos = 0, clusters_to_add;
705 u32 v_cpos = ocfs2_bytes_to_clusters(osb->sb, i_size_read(inode));
706 unsigned int num_clusters = 0;
707 unsigned int ext_flags = 0;
708 u32 size_div, offset_div;
713 u64 s = i_size_read(inode);
715 offset_div = do_div(o, osb->s_clustersize);
716 size_div = do_div(s, osb->s_clustersize);
719 if (offset <= i_size_read(inode))
722 clusters_to_add = ocfs2_bytes_to_clusters(inode->i_sb, offset) -
723 ocfs2_bytes_to_clusters(inode->i_sb, i_size_read(inode));
724 total_zero_len = offset - i_size_read(inode);
726 total_zero_len -= offset_div;
728 /* Allocate clusters to fill out holes, and this is only needed
729 * when we add more than one clusters. Otherwise the cluster will
730 * be allocated during direct IO */
731 if (clusters_to_add > 1) {
732 ret = ocfs2_extend_allocation(inode,
733 OCFS2_I(inode)->ip_clusters,
734 clusters_to_add - 1, 0);
741 while (total_zero_len) {
742 ret = ocfs2_get_clusters(inode, v_cpos, &p_cpos, &num_clusters,
749 zero_start = ocfs2_clusters_to_bytes(osb->sb, p_cpos) +
751 zero_len = ocfs2_clusters_to_bytes(osb->sb, num_clusters) -
753 zero_len = min(total_zero_len, zero_len);
755 if (p_cpos && !(ext_flags & OCFS2_EXT_UNWRITTEN)) {
756 ret = blkdev_issue_zeroout(osb->sb->s_bdev,
757 zero_start >> 9, zero_len >> 9,
765 total_zero_len -= zero_len;
766 v_cpos += ocfs2_bytes_to_clusters(osb->sb, zero_len + size_div);
768 /* Only at first iteration can be cluster not aligned.
769 * So set size_div to 0 for the rest */
777 static ssize_t ocfs2_direct_IO_write(struct kiocb *iocb,
778 struct iov_iter *iter,
783 bool orphaned = false;
784 int is_overwrite = 0;
785 struct file *file = iocb->ki_filp;
786 struct inode *inode = file_inode(file)->i_mapping->host;
787 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
788 struct buffer_head *di_bh = NULL;
789 size_t count = iter->count;
790 journal_t *journal = osb->journal->j_journal;
791 u64 zero_len_head, zero_len_tail;
792 int cluster_align_head, cluster_align_tail;
793 loff_t final_size = offset + count;
794 int append_write = offset >= i_size_read(inode) ? 1 : 0;
795 unsigned int num_clusters = 0;
796 unsigned int ext_flags = 0;
800 u64 s = i_size_read(inode);
802 zero_len_head = do_div(o, 1 << osb->s_clustersize_bits);
803 cluster_align_head = !zero_len_head;
805 zero_len_tail = osb->s_clustersize -
806 do_div(s, osb->s_clustersize);
807 if ((offset - i_size_read(inode)) < zero_len_tail)
808 zero_len_tail = offset - i_size_read(inode);
809 cluster_align_tail = !zero_len_tail;
813 * when final_size > inode->i_size, inode->i_size will be
814 * updated after direct write, so add the inode to orphan
817 if (final_size > i_size_read(inode)) {
818 ret = ocfs2_add_inode_to_orphan(osb, inode);
827 ret = ocfs2_inode_lock(inode, NULL, 1);
833 /* zeroing out the previously allocated cluster tail
834 * that but not zeroed */
835 if (ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)))
836 ret = ocfs2_direct_IO_zero_extend(osb, inode, offset,
837 zero_len_tail, cluster_align_tail);
839 ret = ocfs2_direct_IO_extend_no_holes(osb, inode,
843 ocfs2_inode_unlock(inode, 1);
847 is_overwrite = ocfs2_is_overwrite(osb, inode, offset);
848 if (is_overwrite < 0) {
849 mlog_errno(is_overwrite);
850 ocfs2_inode_unlock(inode, 1);
854 ocfs2_inode_unlock(inode, 1);
857 written = __blockdev_direct_IO(WRITE, iocb, inode, inode->i_sb->s_bdev,
859 ocfs2_direct_IO_get_blocks,
860 ocfs2_dio_end_io, NULL, 0);
861 if (unlikely(written < 0)) {
862 loff_t i_size = i_size_read(inode);
864 if (offset + count > i_size) {
865 ret = ocfs2_inode_lock(inode, &di_bh, 1);
871 if (i_size == i_size_read(inode)) {
872 ret = ocfs2_truncate_file(inode, di_bh,
878 ocfs2_inode_unlock(inode, 1);
884 ocfs2_inode_unlock(inode, 1);
887 ret = jbd2_journal_force_commit(journal);
891 } else if (written > 0 && append_write && !is_overwrite &&
892 !cluster_align_head) {
893 /* zeroing out the allocated cluster head */
895 u32 v_cpos = ocfs2_bytes_to_clusters(osb->sb, offset);
897 ret = ocfs2_inode_lock(inode, NULL, 0);
903 ret = ocfs2_get_clusters(inode, v_cpos, &p_cpos,
904 &num_clusters, &ext_flags);
907 ocfs2_inode_unlock(inode, 0);
911 BUG_ON(!p_cpos || (ext_flags & OCFS2_EXT_UNWRITTEN));
913 ret = blkdev_issue_zeroout(osb->sb->s_bdev,
914 p_cpos << (osb->s_clustersize_bits - 9),
915 zero_len_head >> 9, GFP_NOFS, false);
919 ocfs2_inode_unlock(inode, 0);
925 int update_isize = written > 0 ? 1 : 0;
926 loff_t end = update_isize ? offset + written : 0;
928 tmp_ret = ocfs2_del_inode_from_orphan(osb, inode,
935 tmp_ret = jbd2_journal_force_commit(journal);
948 static ssize_t ocfs2_direct_IO(int rw,
950 struct iov_iter *iter,
953 struct file *file = iocb->ki_filp;
954 struct inode *inode = file_inode(file)->i_mapping->host;
955 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
956 int full_coherency = !(osb->s_mount_opt &
957 OCFS2_MOUNT_COHERENCY_BUFFERED);
960 * Fallback to buffered I/O if we see an inode without
963 if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
966 /* Fallback to buffered I/O if we are appending and
967 * concurrent O_DIRECT writes are allowed.
969 if (i_size_read(inode) <= offset && !full_coherency)
973 return __blockdev_direct_IO(rw, iocb, inode,
976 ocfs2_direct_IO_get_blocks,
977 ocfs2_dio_end_io, NULL, 0);
979 return ocfs2_direct_IO_write(iocb, iter, offset);
982 static void ocfs2_figure_cluster_boundaries(struct ocfs2_super *osb,
987 unsigned int cluster_start = 0, cluster_end = PAGE_CACHE_SIZE;
989 if (unlikely(PAGE_CACHE_SHIFT > osb->s_clustersize_bits)) {
992 cpp = 1 << (PAGE_CACHE_SHIFT - osb->s_clustersize_bits);
994 cluster_start = cpos % cpp;
995 cluster_start = cluster_start << osb->s_clustersize_bits;
997 cluster_end = cluster_start + osb->s_clustersize;
1000 BUG_ON(cluster_start > PAGE_SIZE);
1001 BUG_ON(cluster_end > PAGE_SIZE);
1004 *start = cluster_start;
1010 * 'from' and 'to' are the region in the page to avoid zeroing.
1012 * If pagesize > clustersize, this function will avoid zeroing outside
1013 * of the cluster boundary.
1015 * from == to == 0 is code for "zero the entire cluster region"
1017 static void ocfs2_clear_page_regions(struct page *page,
1018 struct ocfs2_super *osb, u32 cpos,
1019 unsigned from, unsigned to)
1022 unsigned int cluster_start, cluster_end;
1024 ocfs2_figure_cluster_boundaries(osb, cpos, &cluster_start, &cluster_end);
1026 kaddr = kmap_atomic(page);
1029 if (from > cluster_start)
1030 memset(kaddr + cluster_start, 0, from - cluster_start);
1031 if (to < cluster_end)
1032 memset(kaddr + to, 0, cluster_end - to);
1034 memset(kaddr + cluster_start, 0, cluster_end - cluster_start);
1037 kunmap_atomic(kaddr);
1041 * Nonsparse file systems fully allocate before we get to the write
1042 * code. This prevents ocfs2_write() from tagging the write as an
1043 * allocating one, which means ocfs2_map_page_blocks() might try to
1044 * read-in the blocks at the tail of our file. Avoid reading them by
1045 * testing i_size against each block offset.
1047 static int ocfs2_should_read_blk(struct inode *inode, struct page *page,
1048 unsigned int block_start)
1050 u64 offset = page_offset(page) + block_start;
1052 if (ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)))
1055 if (i_size_read(inode) > offset)
1062 * Some of this taken from __block_write_begin(). We already have our
1063 * mapping by now though, and the entire write will be allocating or
1064 * it won't, so not much need to use BH_New.
1066 * This will also skip zeroing, which is handled externally.
1068 int ocfs2_map_page_blocks(struct page *page, u64 *p_blkno,
1069 struct inode *inode, unsigned int from,
1070 unsigned int to, int new)
1073 struct buffer_head *head, *bh, *wait[2], **wait_bh = wait;
1074 unsigned int block_end, block_start;
1075 unsigned int bsize = 1 << inode->i_blkbits;
1077 if (!page_has_buffers(page))
1078 create_empty_buffers(page, bsize, 0);
1080 head = page_buffers(page);
1081 for (bh = head, block_start = 0; bh != head || !block_start;
1082 bh = bh->b_this_page, block_start += bsize) {
1083 block_end = block_start + bsize;
1085 clear_buffer_new(bh);
1088 * Ignore blocks outside of our i/o range -
1089 * they may belong to unallocated clusters.
1091 if (block_start >= to || block_end <= from) {
1092 if (PageUptodate(page))
1093 set_buffer_uptodate(bh);
1098 * For an allocating write with cluster size >= page
1099 * size, we always write the entire page.
1104 if (!buffer_mapped(bh)) {
1105 map_bh(bh, inode->i_sb, *p_blkno);
1106 unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
1109 if (PageUptodate(page)) {
1110 if (!buffer_uptodate(bh))
1111 set_buffer_uptodate(bh);
1112 } else if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
1114 ocfs2_should_read_blk(inode, page, block_start) &&
1115 (block_start < from || block_end > to)) {
1116 ll_rw_block(READ, 1, &bh);
1120 *p_blkno = *p_blkno + 1;
1124 * If we issued read requests - let them complete.
1126 while(wait_bh > wait) {
1127 wait_on_buffer(*--wait_bh);
1128 if (!buffer_uptodate(*wait_bh))
1132 if (ret == 0 || !new)
1136 * If we get -EIO above, zero out any newly allocated blocks
1137 * to avoid exposing stale data.
1142 block_end = block_start + bsize;
1143 if (block_end <= from)
1145 if (block_start >= to)
1148 zero_user(page, block_start, bh->b_size);
1149 set_buffer_uptodate(bh);
1150 mark_buffer_dirty(bh);
1153 block_start = block_end;
1154 bh = bh->b_this_page;
1155 } while (bh != head);
1160 #if (PAGE_CACHE_SIZE >= OCFS2_MAX_CLUSTERSIZE)
1161 #define OCFS2_MAX_CTXT_PAGES 1
1163 #define OCFS2_MAX_CTXT_PAGES (OCFS2_MAX_CLUSTERSIZE / PAGE_CACHE_SIZE)
1166 #define OCFS2_MAX_CLUSTERS_PER_PAGE (PAGE_CACHE_SIZE / OCFS2_MIN_CLUSTERSIZE)
1169 * Describe the state of a single cluster to be written to.
1171 struct ocfs2_write_cluster_desc {
1175 * Give this a unique field because c_phys eventually gets
1179 unsigned c_unwritten;
1180 unsigned c_needs_zero;
1183 struct ocfs2_write_ctxt {
1184 /* Logical cluster position / len of write */
1188 /* First cluster allocated in a nonsparse extend */
1189 u32 w_first_new_cpos;
1191 struct ocfs2_write_cluster_desc w_desc[OCFS2_MAX_CLUSTERS_PER_PAGE];
1194 * This is true if page_size > cluster_size.
1196 * It triggers a set of special cases during write which might
1197 * have to deal with allocating writes to partial pages.
1199 unsigned int w_large_pages;
1202 * Pages involved in this write.
1204 * w_target_page is the page being written to by the user.
1206 * w_pages is an array of pages which always contains
1207 * w_target_page, and in the case of an allocating write with
1208 * page_size < cluster size, it will contain zero'd and mapped
1209 * pages adjacent to w_target_page which need to be written
1210 * out in so that future reads from that region will get
1213 unsigned int w_num_pages;
1214 struct page *w_pages[OCFS2_MAX_CTXT_PAGES];
1215 struct page *w_target_page;
1218 * w_target_locked is used for page_mkwrite path indicating no unlocking
1219 * against w_target_page in ocfs2_write_end_nolock.
1221 unsigned int w_target_locked:1;
1224 * ocfs2_write_end() uses this to know what the real range to
1225 * write in the target should be.
1227 unsigned int w_target_from;
1228 unsigned int w_target_to;
1231 * We could use journal_current_handle() but this is cleaner,
1236 struct buffer_head *w_di_bh;
1238 struct ocfs2_cached_dealloc_ctxt w_dealloc;
1241 void ocfs2_unlock_and_free_pages(struct page **pages, int num_pages)
1245 for(i = 0; i < num_pages; i++) {
1247 unlock_page(pages[i]);
1248 mark_page_accessed(pages[i]);
1249 page_cache_release(pages[i]);
1254 static void ocfs2_unlock_pages(struct ocfs2_write_ctxt *wc)
1259 * w_target_locked is only set to true in the page_mkwrite() case.
1260 * The intent is to allow us to lock the target page from write_begin()
1261 * to write_end(). The caller must hold a ref on w_target_page.
1263 if (wc->w_target_locked) {
1264 BUG_ON(!wc->w_target_page);
1265 for (i = 0; i < wc->w_num_pages; i++) {
1266 if (wc->w_target_page == wc->w_pages[i]) {
1267 wc->w_pages[i] = NULL;
1271 mark_page_accessed(wc->w_target_page);
1272 page_cache_release(wc->w_target_page);
1274 ocfs2_unlock_and_free_pages(wc->w_pages, wc->w_num_pages);
1277 static void ocfs2_free_write_ctxt(struct ocfs2_write_ctxt *wc)
1279 ocfs2_unlock_pages(wc);
1280 brelse(wc->w_di_bh);
1284 static int ocfs2_alloc_write_ctxt(struct ocfs2_write_ctxt **wcp,
1285 struct ocfs2_super *osb, loff_t pos,
1286 unsigned len, struct buffer_head *di_bh)
1289 struct ocfs2_write_ctxt *wc;
1291 wc = kzalloc(sizeof(struct ocfs2_write_ctxt), GFP_NOFS);
1295 wc->w_cpos = pos >> osb->s_clustersize_bits;
1296 wc->w_first_new_cpos = UINT_MAX;
1297 cend = (pos + len - 1) >> osb->s_clustersize_bits;
1298 wc->w_clen = cend - wc->w_cpos + 1;
1300 wc->w_di_bh = di_bh;
1302 if (unlikely(PAGE_CACHE_SHIFT > osb->s_clustersize_bits))
1303 wc->w_large_pages = 1;
1305 wc->w_large_pages = 0;
1307 ocfs2_init_dealloc_ctxt(&wc->w_dealloc);
1315 * If a page has any new buffers, zero them out here, and mark them uptodate
1316 * and dirty so they'll be written out (in order to prevent uninitialised
1317 * block data from leaking). And clear the new bit.
1319 static void ocfs2_zero_new_buffers(struct page *page, unsigned from, unsigned to)
1321 unsigned int block_start, block_end;
1322 struct buffer_head *head, *bh;
1324 BUG_ON(!PageLocked(page));
1325 if (!page_has_buffers(page))
1328 bh = head = page_buffers(page);
1331 block_end = block_start + bh->b_size;
1333 if (buffer_new(bh)) {
1334 if (block_end > from && block_start < to) {
1335 if (!PageUptodate(page)) {
1336 unsigned start, end;
1338 start = max(from, block_start);
1339 end = min(to, block_end);
1341 zero_user_segment(page, start, end);
1342 set_buffer_uptodate(bh);
1345 clear_buffer_new(bh);
1346 mark_buffer_dirty(bh);
1350 block_start = block_end;
1351 bh = bh->b_this_page;
1352 } while (bh != head);
1356 * Only called when we have a failure during allocating write to write
1357 * zero's to the newly allocated region.
1359 static void ocfs2_write_failure(struct inode *inode,
1360 struct ocfs2_write_ctxt *wc,
1361 loff_t user_pos, unsigned user_len)
1364 unsigned from = user_pos & (PAGE_CACHE_SIZE - 1),
1365 to = user_pos + user_len;
1366 struct page *tmppage;
1368 ocfs2_zero_new_buffers(wc->w_target_page, from, to);
1370 for(i = 0; i < wc->w_num_pages; i++) {
1371 tmppage = wc->w_pages[i];
1373 if (page_has_buffers(tmppage)) {
1374 if (ocfs2_should_order_data(inode))
1375 ocfs2_jbd2_file_inode(wc->w_handle, inode);
1377 block_commit_write(tmppage, from, to);
1382 static int ocfs2_prepare_page_for_write(struct inode *inode, u64 *p_blkno,
1383 struct ocfs2_write_ctxt *wc,
1384 struct page *page, u32 cpos,
1385 loff_t user_pos, unsigned user_len,
1389 unsigned int map_from = 0, map_to = 0;
1390 unsigned int cluster_start, cluster_end;
1391 unsigned int user_data_from = 0, user_data_to = 0;
1393 ocfs2_figure_cluster_boundaries(OCFS2_SB(inode->i_sb), cpos,
1394 &cluster_start, &cluster_end);
1396 /* treat the write as new if the a hole/lseek spanned across
1397 * the page boundary.
1399 new = new | ((i_size_read(inode) <= page_offset(page)) &&
1400 (page_offset(page) <= user_pos));
1402 if (page == wc->w_target_page) {
1403 map_from = user_pos & (PAGE_CACHE_SIZE - 1);
1404 map_to = map_from + user_len;
1407 ret = ocfs2_map_page_blocks(page, p_blkno, inode,
1408 cluster_start, cluster_end,
1411 ret = ocfs2_map_page_blocks(page, p_blkno, inode,
1412 map_from, map_to, new);
1418 user_data_from = map_from;
1419 user_data_to = map_to;
1421 map_from = cluster_start;
1422 map_to = cluster_end;
1426 * If we haven't allocated the new page yet, we
1427 * shouldn't be writing it out without copying user
1428 * data. This is likely a math error from the caller.
1432 map_from = cluster_start;
1433 map_to = cluster_end;
1435 ret = ocfs2_map_page_blocks(page, p_blkno, inode,
1436 cluster_start, cluster_end, new);
1444 * Parts of newly allocated pages need to be zero'd.
1446 * Above, we have also rewritten 'to' and 'from' - as far as
1447 * the rest of the function is concerned, the entire cluster
1448 * range inside of a page needs to be written.
1450 * We can skip this if the page is up to date - it's already
1451 * been zero'd from being read in as a hole.
1453 if (new && !PageUptodate(page))
1454 ocfs2_clear_page_regions(page, OCFS2_SB(inode->i_sb),
1455 cpos, user_data_from, user_data_to);
1457 flush_dcache_page(page);
1464 * This function will only grab one clusters worth of pages.
1466 static int ocfs2_grab_pages_for_write(struct address_space *mapping,
1467 struct ocfs2_write_ctxt *wc,
1468 u32 cpos, loff_t user_pos,
1469 unsigned user_len, int new,
1470 struct page *mmap_page)
1473 unsigned long start, target_index, end_index, index;
1474 struct inode *inode = mapping->host;
1477 target_index = user_pos >> PAGE_CACHE_SHIFT;
1480 * Figure out how many pages we'll be manipulating here. For
1481 * non allocating write, we just change the one
1482 * page. Otherwise, we'll need a whole clusters worth. If we're
1483 * writing past i_size, we only need enough pages to cover the
1484 * last page of the write.
1487 wc->w_num_pages = ocfs2_pages_per_cluster(inode->i_sb);
1488 start = ocfs2_align_clusters_to_page_index(inode->i_sb, cpos);
1490 * We need the index *past* the last page we could possibly
1491 * touch. This is the page past the end of the write or
1492 * i_size, whichever is greater.
1494 last_byte = max(user_pos + user_len, i_size_read(inode));
1495 BUG_ON(last_byte < 1);
1496 end_index = ((last_byte - 1) >> PAGE_CACHE_SHIFT) + 1;
1497 if ((start + wc->w_num_pages) > end_index)
1498 wc->w_num_pages = end_index - start;
1500 wc->w_num_pages = 1;
1501 start = target_index;
1504 for(i = 0; i < wc->w_num_pages; i++) {
1507 if (index == target_index && mmap_page) {
1509 * ocfs2_pagemkwrite() is a little different
1510 * and wants us to directly use the page
1513 lock_page(mmap_page);
1515 /* Exit and let the caller retry */
1516 if (mmap_page->mapping != mapping) {
1517 WARN_ON(mmap_page->mapping);
1518 unlock_page(mmap_page);
1523 page_cache_get(mmap_page);
1524 wc->w_pages[i] = mmap_page;
1525 wc->w_target_locked = true;
1527 wc->w_pages[i] = find_or_create_page(mapping, index,
1529 if (!wc->w_pages[i]) {
1535 wait_for_stable_page(wc->w_pages[i]);
1537 if (index == target_index)
1538 wc->w_target_page = wc->w_pages[i];
1542 wc->w_target_locked = false;
1547 * Prepare a single cluster for write one cluster into the file.
1549 static int ocfs2_write_cluster(struct address_space *mapping,
1550 u32 phys, unsigned int unwritten,
1551 unsigned int should_zero,
1552 struct ocfs2_alloc_context *data_ac,
1553 struct ocfs2_alloc_context *meta_ac,
1554 struct ocfs2_write_ctxt *wc, u32 cpos,
1555 loff_t user_pos, unsigned user_len)
1558 u64 v_blkno, p_blkno;
1559 struct inode *inode = mapping->host;
1560 struct ocfs2_extent_tree et;
1562 new = phys == 0 ? 1 : 0;
1567 * This is safe to call with the page locks - it won't take
1568 * any additional semaphores or cluster locks.
1571 ret = ocfs2_add_inode_data(OCFS2_SB(inode->i_sb), inode,
1572 &tmp_pos, 1, 0, wc->w_di_bh,
1573 wc->w_handle, data_ac,
1576 * This shouldn't happen because we must have already
1577 * calculated the correct meta data allocation required. The
1578 * internal tree allocation code should know how to increase
1579 * transaction credits itself.
1581 * If need be, we could handle -EAGAIN for a
1582 * RESTART_TRANS here.
1584 mlog_bug_on_msg(ret == -EAGAIN,
1585 "Inode %llu: EAGAIN return during allocation.\n",
1586 (unsigned long long)OCFS2_I(inode)->ip_blkno);
1591 } else if (unwritten) {
1592 ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode),
1594 ret = ocfs2_mark_extent_written(inode, &et,
1595 wc->w_handle, cpos, 1, phys,
1596 meta_ac, &wc->w_dealloc);
1604 v_blkno = ocfs2_clusters_to_blocks(inode->i_sb, cpos);
1606 v_blkno = user_pos >> inode->i_sb->s_blocksize_bits;
1609 * The only reason this should fail is due to an inability to
1610 * find the extent added.
1612 ret = ocfs2_extent_map_get_blocks(inode, v_blkno, &p_blkno, NULL,
1615 mlog(ML_ERROR, "Get physical blkno failed for inode %llu, "
1616 "at logical block %llu",
1617 (unsigned long long)OCFS2_I(inode)->ip_blkno,
1618 (unsigned long long)v_blkno);
1622 BUG_ON(p_blkno == 0);
1624 for(i = 0; i < wc->w_num_pages; i++) {
1627 tmpret = ocfs2_prepare_page_for_write(inode, &p_blkno, wc,
1628 wc->w_pages[i], cpos,
1639 * We only have cleanup to do in case of allocating write.
1642 ocfs2_write_failure(inode, wc, user_pos, user_len);
1649 static int ocfs2_write_cluster_by_desc(struct address_space *mapping,
1650 struct ocfs2_alloc_context *data_ac,
1651 struct ocfs2_alloc_context *meta_ac,
1652 struct ocfs2_write_ctxt *wc,
1653 loff_t pos, unsigned len)
1657 unsigned int local_len = len;
1658 struct ocfs2_write_cluster_desc *desc;
1659 struct ocfs2_super *osb = OCFS2_SB(mapping->host->i_sb);
1661 for (i = 0; i < wc->w_clen; i++) {
1662 desc = &wc->w_desc[i];
1665 * We have to make sure that the total write passed in
1666 * doesn't extend past a single cluster.
1669 cluster_off = pos & (osb->s_clustersize - 1);
1670 if ((cluster_off + local_len) > osb->s_clustersize)
1671 local_len = osb->s_clustersize - cluster_off;
1673 ret = ocfs2_write_cluster(mapping, desc->c_phys,
1677 wc, desc->c_cpos, pos, local_len);
1693 * ocfs2_write_end() wants to know which parts of the target page it
1694 * should complete the write on. It's easiest to compute them ahead of
1695 * time when a more complete view of the write is available.
1697 static void ocfs2_set_target_boundaries(struct ocfs2_super *osb,
1698 struct ocfs2_write_ctxt *wc,
1699 loff_t pos, unsigned len, int alloc)
1701 struct ocfs2_write_cluster_desc *desc;
1703 wc->w_target_from = pos & (PAGE_CACHE_SIZE - 1);
1704 wc->w_target_to = wc->w_target_from + len;
1710 * Allocating write - we may have different boundaries based
1711 * on page size and cluster size.
1713 * NOTE: We can no longer compute one value from the other as
1714 * the actual write length and user provided length may be
1718 if (wc->w_large_pages) {
1720 * We only care about the 1st and last cluster within
1721 * our range and whether they should be zero'd or not. Either
1722 * value may be extended out to the start/end of a
1723 * newly allocated cluster.
1725 desc = &wc->w_desc[0];
1726 if (desc->c_needs_zero)
1727 ocfs2_figure_cluster_boundaries(osb,
1732 desc = &wc->w_desc[wc->w_clen - 1];
1733 if (desc->c_needs_zero)
1734 ocfs2_figure_cluster_boundaries(osb,
1739 wc->w_target_from = 0;
1740 wc->w_target_to = PAGE_CACHE_SIZE;
1745 * Populate each single-cluster write descriptor in the write context
1746 * with information about the i/o to be done.
1748 * Returns the number of clusters that will have to be allocated, as
1749 * well as a worst case estimate of the number of extent records that
1750 * would have to be created during a write to an unwritten region.
1752 static int ocfs2_populate_write_desc(struct inode *inode,
1753 struct ocfs2_write_ctxt *wc,
1754 unsigned int *clusters_to_alloc,
1755 unsigned int *extents_to_split)
1758 struct ocfs2_write_cluster_desc *desc;
1759 unsigned int num_clusters = 0;
1760 unsigned int ext_flags = 0;
1764 *clusters_to_alloc = 0;
1765 *extents_to_split = 0;
1767 for (i = 0; i < wc->w_clen; i++) {
1768 desc = &wc->w_desc[i];
1769 desc->c_cpos = wc->w_cpos + i;
1771 if (num_clusters == 0) {
1773 * Need to look up the next extent record.
1775 ret = ocfs2_get_clusters(inode, desc->c_cpos, &phys,
1776 &num_clusters, &ext_flags);
1782 /* We should already CoW the refcountd extent. */
1783 BUG_ON(ext_flags & OCFS2_EXT_REFCOUNTED);
1786 * Assume worst case - that we're writing in
1787 * the middle of the extent.
1789 * We can assume that the write proceeds from
1790 * left to right, in which case the extent
1791 * insert code is smart enough to coalesce the
1792 * next splits into the previous records created.
1794 if (ext_flags & OCFS2_EXT_UNWRITTEN)
1795 *extents_to_split = *extents_to_split + 2;
1798 * Only increment phys if it doesn't describe
1805 * If w_first_new_cpos is < UINT_MAX, we have a non-sparse
1806 * file that got extended. w_first_new_cpos tells us
1807 * where the newly allocated clusters are so we can
1810 if (desc->c_cpos >= wc->w_first_new_cpos) {
1812 desc->c_needs_zero = 1;
1815 desc->c_phys = phys;
1818 desc->c_needs_zero = 1;
1819 *clusters_to_alloc = *clusters_to_alloc + 1;
1822 if (ext_flags & OCFS2_EXT_UNWRITTEN) {
1823 desc->c_unwritten = 1;
1824 desc->c_needs_zero = 1;
1835 static int ocfs2_write_begin_inline(struct address_space *mapping,
1836 struct inode *inode,
1837 struct ocfs2_write_ctxt *wc)
1840 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1843 struct ocfs2_dinode *di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1845 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1846 if (IS_ERR(handle)) {
1847 ret = PTR_ERR(handle);
1852 page = find_or_create_page(mapping, 0, GFP_NOFS);
1854 ocfs2_commit_trans(osb, handle);
1860 * If we don't set w_num_pages then this page won't get unlocked
1861 * and freed on cleanup of the write context.
1863 wc->w_pages[0] = wc->w_target_page = page;
1864 wc->w_num_pages = 1;
1866 ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), wc->w_di_bh,
1867 OCFS2_JOURNAL_ACCESS_WRITE);
1869 ocfs2_commit_trans(osb, handle);
1875 if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL))
1876 ocfs2_set_inode_data_inline(inode, di);
1878 if (!PageUptodate(page)) {
1879 ret = ocfs2_read_inline_data(inode, page, wc->w_di_bh);
1881 ocfs2_commit_trans(osb, handle);
1887 wc->w_handle = handle;
1892 int ocfs2_size_fits_inline_data(struct buffer_head *di_bh, u64 new_size)
1894 struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
1896 if (new_size <= le16_to_cpu(di->id2.i_data.id_count))
1901 static int ocfs2_try_to_write_inline_data(struct address_space *mapping,
1902 struct inode *inode, loff_t pos,
1903 unsigned len, struct page *mmap_page,
1904 struct ocfs2_write_ctxt *wc)
1906 int ret, written = 0;
1907 loff_t end = pos + len;
1908 struct ocfs2_inode_info *oi = OCFS2_I(inode);
1909 struct ocfs2_dinode *di = NULL;
1911 trace_ocfs2_try_to_write_inline_data((unsigned long long)oi->ip_blkno,
1912 len, (unsigned long long)pos,
1913 oi->ip_dyn_features);
1916 * Handle inodes which already have inline data 1st.
1918 if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1919 if (mmap_page == NULL &&
1920 ocfs2_size_fits_inline_data(wc->w_di_bh, end))
1921 goto do_inline_write;
1924 * The write won't fit - we have to give this inode an
1925 * inline extent list now.
1927 ret = ocfs2_convert_inline_data_to_extents(inode, wc->w_di_bh);
1934 * Check whether the inode can accept inline data.
1936 if (oi->ip_clusters != 0 || i_size_read(inode) != 0)
1940 * Check whether the write can fit.
1942 di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1944 end > ocfs2_max_inline_data_with_xattr(inode->i_sb, di))
1948 ret = ocfs2_write_begin_inline(mapping, inode, wc);
1955 * This signals to the caller that the data can be written
1960 return written ? written : ret;
1964 * This function only does anything for file systems which can't
1965 * handle sparse files.
1967 * What we want to do here is fill in any hole between the current end
1968 * of allocation and the end of our write. That way the rest of the
1969 * write path can treat it as an non-allocating write, which has no
1970 * special case code for sparse/nonsparse files.
1972 static int ocfs2_expand_nonsparse_inode(struct inode *inode,
1973 struct buffer_head *di_bh,
1974 loff_t pos, unsigned len,
1975 struct ocfs2_write_ctxt *wc)
1978 loff_t newsize = pos + len;
1980 BUG_ON(ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)));
1982 if (newsize <= i_size_read(inode))
1985 ret = ocfs2_extend_no_holes(inode, di_bh, newsize, pos);
1989 wc->w_first_new_cpos =
1990 ocfs2_clusters_for_bytes(inode->i_sb, i_size_read(inode));
1995 static int ocfs2_zero_tail(struct inode *inode, struct buffer_head *di_bh,
2000 BUG_ON(!ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)));
2001 if (pos > i_size_read(inode))
2002 ret = ocfs2_zero_extend(inode, di_bh, pos);
2008 * Try to flush truncate logs if we can free enough clusters from it.
2009 * As for return value, "< 0" means error, "0" no space and "1" means
2010 * we have freed enough spaces and let the caller try to allocate again.
2012 static int ocfs2_try_to_free_truncate_log(struct ocfs2_super *osb,
2013 unsigned int needed)
2017 unsigned int truncated_clusters;
2019 mutex_lock(&osb->osb_tl_inode->i_mutex);
2020 truncated_clusters = osb->truncated_clusters;
2021 mutex_unlock(&osb->osb_tl_inode->i_mutex);
2024 * Check whether we can succeed in allocating if we free
2027 if (truncated_clusters < needed)
2030 ret = ocfs2_flush_truncate_log(osb);
2036 if (jbd2_journal_start_commit(osb->journal->j_journal, &target)) {
2037 jbd2_log_wait_commit(osb->journal->j_journal, target);
2044 int ocfs2_write_begin_nolock(struct file *filp,
2045 struct address_space *mapping,
2046 loff_t pos, unsigned len, unsigned flags,
2047 struct page **pagep, void **fsdata,
2048 struct buffer_head *di_bh, struct page *mmap_page)
2050 int ret, cluster_of_pages, credits = OCFS2_INODE_UPDATE_CREDITS;
2051 unsigned int clusters_to_alloc, extents_to_split, clusters_need = 0;
2052 struct ocfs2_write_ctxt *wc;
2053 struct inode *inode = mapping->host;
2054 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
2055 struct ocfs2_dinode *di;
2056 struct ocfs2_alloc_context *data_ac = NULL;
2057 struct ocfs2_alloc_context *meta_ac = NULL;
2059 struct ocfs2_extent_tree et;
2060 int try_free = 1, ret1;
2063 ret = ocfs2_alloc_write_ctxt(&wc, osb, pos, len, di_bh);
2069 if (ocfs2_supports_inline_data(osb)) {
2070 ret = ocfs2_try_to_write_inline_data(mapping, inode, pos, len,
2082 if (ocfs2_sparse_alloc(osb))
2083 ret = ocfs2_zero_tail(inode, di_bh, pos);
2085 ret = ocfs2_expand_nonsparse_inode(inode, di_bh, pos, len,
2092 ret = ocfs2_check_range_for_refcount(inode, pos, len);
2096 } else if (ret == 1) {
2097 clusters_need = wc->w_clen;
2098 ret = ocfs2_refcount_cow(inode, di_bh,
2099 wc->w_cpos, wc->w_clen, UINT_MAX);
2106 ret = ocfs2_populate_write_desc(inode, wc, &clusters_to_alloc,
2112 clusters_need += clusters_to_alloc;
2114 di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
2116 trace_ocfs2_write_begin_nolock(
2117 (unsigned long long)OCFS2_I(inode)->ip_blkno,
2118 (long long)i_size_read(inode),
2119 le32_to_cpu(di->i_clusters),
2120 pos, len, flags, mmap_page,
2121 clusters_to_alloc, extents_to_split);
2124 * We set w_target_from, w_target_to here so that
2125 * ocfs2_write_end() knows which range in the target page to
2126 * write out. An allocation requires that we write the entire
2129 if (clusters_to_alloc || extents_to_split) {
2131 * XXX: We are stretching the limits of
2132 * ocfs2_lock_allocators(). It greatly over-estimates
2133 * the work to be done.
2135 ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode),
2137 ret = ocfs2_lock_allocators(inode, &et,
2138 clusters_to_alloc, extents_to_split,
2139 &data_ac, &meta_ac);
2146 data_ac->ac_resv = &OCFS2_I(inode)->ip_la_data_resv;
2148 credits = ocfs2_calc_extend_credits(inode->i_sb,
2154 * We have to zero sparse allocated clusters, unwritten extent clusters,
2155 * and non-sparse clusters we just extended. For non-sparse writes,
2156 * we know zeros will only be needed in the first and/or last cluster.
2158 if (clusters_to_alloc || extents_to_split ||
2159 (wc->w_clen && (wc->w_desc[0].c_needs_zero ||
2160 wc->w_desc[wc->w_clen - 1].c_needs_zero)))
2161 cluster_of_pages = 1;
2163 cluster_of_pages = 0;
2165 ocfs2_set_target_boundaries(osb, wc, pos, len, cluster_of_pages);
2167 handle = ocfs2_start_trans(osb, credits);
2168 if (IS_ERR(handle)) {
2169 ret = PTR_ERR(handle);
2174 wc->w_handle = handle;
2176 if (clusters_to_alloc) {
2177 ret = dquot_alloc_space_nodirty(inode,
2178 ocfs2_clusters_to_bytes(osb->sb, clusters_to_alloc));
2183 * We don't want this to fail in ocfs2_write_end(), so do it
2186 ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), wc->w_di_bh,
2187 OCFS2_JOURNAL_ACCESS_WRITE);
2194 * Fill our page array first. That way we've grabbed enough so
2195 * that we can zero and flush if we error after adding the
2198 ret = ocfs2_grab_pages_for_write(mapping, wc, wc->w_cpos, pos, len,
2199 cluster_of_pages, mmap_page);
2200 if (ret && ret != -EAGAIN) {
2206 * ocfs2_grab_pages_for_write() returns -EAGAIN if it could not lock
2207 * the target page. In this case, we exit with no error and no target
2208 * page. This will trigger the caller, page_mkwrite(), to re-try
2211 if (ret == -EAGAIN) {
2212 BUG_ON(wc->w_target_page);
2217 ret = ocfs2_write_cluster_by_desc(mapping, data_ac, meta_ac, wc, pos,
2225 ocfs2_free_alloc_context(data_ac);
2227 ocfs2_free_alloc_context(meta_ac);
2230 *pagep = wc->w_target_page;
2234 if (clusters_to_alloc)
2235 dquot_free_space(inode,
2236 ocfs2_clusters_to_bytes(osb->sb, clusters_to_alloc));
2238 ocfs2_commit_trans(osb, handle);
2241 ocfs2_free_write_ctxt(wc);
2244 ocfs2_free_alloc_context(data_ac);
2248 ocfs2_free_alloc_context(meta_ac);
2252 if (ret == -ENOSPC && try_free) {
2254 * Try to free some truncate log so that we can have enough
2255 * clusters to allocate.
2259 ret1 = ocfs2_try_to_free_truncate_log(osb, clusters_need);
2270 static int ocfs2_write_begin(struct file *file, struct address_space *mapping,
2271 loff_t pos, unsigned len, unsigned flags,
2272 struct page **pagep, void **fsdata)
2275 struct buffer_head *di_bh = NULL;
2276 struct inode *inode = mapping->host;
2278 ret = ocfs2_inode_lock(inode, &di_bh, 1);
2285 * Take alloc sem here to prevent concurrent lookups. That way
2286 * the mapping, zeroing and tree manipulation within
2287 * ocfs2_write() will be safe against ->readpage(). This
2288 * should also serve to lock out allocation from a shared
2291 down_write(&OCFS2_I(inode)->ip_alloc_sem);
2293 ret = ocfs2_write_begin_nolock(file, mapping, pos, len, flags, pagep,
2294 fsdata, di_bh, NULL);
2305 up_write(&OCFS2_I(inode)->ip_alloc_sem);
2308 ocfs2_inode_unlock(inode, 1);
2313 static void ocfs2_write_end_inline(struct inode *inode, loff_t pos,
2314 unsigned len, unsigned *copied,
2315 struct ocfs2_dinode *di,
2316 struct ocfs2_write_ctxt *wc)
2320 if (unlikely(*copied < len)) {
2321 if (!PageUptodate(wc->w_target_page)) {
2327 kaddr = kmap_atomic(wc->w_target_page);
2328 memcpy(di->id2.i_data.id_data + pos, kaddr + pos, *copied);
2329 kunmap_atomic(kaddr);
2331 trace_ocfs2_write_end_inline(
2332 (unsigned long long)OCFS2_I(inode)->ip_blkno,
2333 (unsigned long long)pos, *copied,
2334 le16_to_cpu(di->id2.i_data.id_count),
2335 le16_to_cpu(di->i_dyn_features));
2338 int ocfs2_write_end_nolock(struct address_space *mapping,
2339 loff_t pos, unsigned len, unsigned copied,
2340 struct page *page, void *fsdata)
2343 unsigned from, to, start = pos & (PAGE_CACHE_SIZE - 1);
2344 struct inode *inode = mapping->host;
2345 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
2346 struct ocfs2_write_ctxt *wc = fsdata;
2347 struct ocfs2_dinode *di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
2348 handle_t *handle = wc->w_handle;
2349 struct page *tmppage;
2351 if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
2352 ocfs2_write_end_inline(inode, pos, len, &copied, di, wc);
2353 goto out_write_size;
2356 if (unlikely(copied < len)) {
2357 if (!PageUptodate(wc->w_target_page))
2360 ocfs2_zero_new_buffers(wc->w_target_page, start+copied,
2363 flush_dcache_page(wc->w_target_page);
2365 for(i = 0; i < wc->w_num_pages; i++) {
2366 tmppage = wc->w_pages[i];
2368 if (tmppage == wc->w_target_page) {
2369 from = wc->w_target_from;
2370 to = wc->w_target_to;
2372 BUG_ON(from > PAGE_CACHE_SIZE ||
2373 to > PAGE_CACHE_SIZE ||
2377 * Pages adjacent to the target (if any) imply
2378 * a hole-filling write in which case we want
2379 * to flush their entire range.
2382 to = PAGE_CACHE_SIZE;
2385 if (page_has_buffers(tmppage)) {
2386 if (ocfs2_should_order_data(inode))
2387 ocfs2_jbd2_file_inode(wc->w_handle, inode);
2388 block_commit_write(tmppage, from, to);
2394 if (pos > i_size_read(inode)) {
2395 i_size_write(inode, pos);
2396 mark_inode_dirty(inode);
2398 inode->i_blocks = ocfs2_inode_sector_count(inode);
2399 di->i_size = cpu_to_le64((u64)i_size_read(inode));
2400 inode->i_mtime = inode->i_ctime = CURRENT_TIME;
2401 di->i_mtime = di->i_ctime = cpu_to_le64(inode->i_mtime.tv_sec);
2402 di->i_mtime_nsec = di->i_ctime_nsec = cpu_to_le32(inode->i_mtime.tv_nsec);
2403 ocfs2_update_inode_fsync_trans(handle, inode, 1);
2404 ocfs2_journal_dirty(handle, wc->w_di_bh);
2406 /* unlock pages before dealloc since it needs acquiring j_trans_barrier
2407 * lock, or it will cause a deadlock since journal commit threads holds
2408 * this lock and will ask for the page lock when flushing the data.
2409 * put it here to preserve the unlock order.
2411 ocfs2_unlock_pages(wc);
2413 ocfs2_commit_trans(osb, handle);
2415 ocfs2_run_deallocs(osb, &wc->w_dealloc);
2417 brelse(wc->w_di_bh);
2423 static int ocfs2_write_end(struct file *file, struct address_space *mapping,
2424 loff_t pos, unsigned len, unsigned copied,
2425 struct page *page, void *fsdata)
2428 struct inode *inode = mapping->host;
2430 ret = ocfs2_write_end_nolock(mapping, pos, len, copied, page, fsdata);
2432 up_write(&OCFS2_I(inode)->ip_alloc_sem);
2433 ocfs2_inode_unlock(inode, 1);
2438 const struct address_space_operations ocfs2_aops = {
2439 .readpage = ocfs2_readpage,
2440 .readpages = ocfs2_readpages,
2441 .writepage = ocfs2_writepage,
2442 .write_begin = ocfs2_write_begin,
2443 .write_end = ocfs2_write_end,
2445 .direct_IO = ocfs2_direct_IO,
2446 .invalidatepage = block_invalidatepage,
2447 .releasepage = ocfs2_releasepage,
2448 .migratepage = buffer_migrate_page,
2449 .is_partially_uptodate = block_is_partially_uptodate,
2450 .error_remove_page = generic_error_remove_page,