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>
32 #include <linux/uio.h>
34 #include <cluster/masklog.h>
41 #include "extent_map.h"
48 #include "refcounttree.h"
49 #include "ocfs2_trace.h"
51 #include "buffer_head_io.h"
56 static int ocfs2_symlink_get_block(struct inode *inode, sector_t iblock,
57 struct buffer_head *bh_result, int create)
61 struct ocfs2_dinode *fe = NULL;
62 struct buffer_head *bh = NULL;
63 struct buffer_head *buffer_cache_bh = NULL;
64 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
67 trace_ocfs2_symlink_get_block(
68 (unsigned long long)OCFS2_I(inode)->ip_blkno,
69 (unsigned long long)iblock, bh_result, create);
71 BUG_ON(ocfs2_inode_is_fast_symlink(inode));
73 if ((iblock << inode->i_sb->s_blocksize_bits) > PATH_MAX + 1) {
74 mlog(ML_ERROR, "block offset > PATH_MAX: %llu",
75 (unsigned long long)iblock);
79 status = ocfs2_read_inode_block(inode, &bh);
84 fe = (struct ocfs2_dinode *) bh->b_data;
86 if ((u64)iblock >= ocfs2_clusters_to_blocks(inode->i_sb,
87 le32_to_cpu(fe->i_clusters))) {
89 mlog(ML_ERROR, "block offset is outside the allocated size: "
90 "%llu\n", (unsigned long long)iblock);
94 /* We don't use the page cache to create symlink data, so if
95 * need be, copy it over from the buffer cache. */
96 if (!buffer_uptodate(bh_result) && ocfs2_inode_is_new(inode)) {
97 u64 blkno = le64_to_cpu(fe->id2.i_list.l_recs[0].e_blkno) +
99 buffer_cache_bh = sb_getblk(osb->sb, blkno);
100 if (!buffer_cache_bh) {
102 mlog(ML_ERROR, "couldn't getblock for symlink!\n");
106 /* we haven't locked out transactions, so a commit
107 * could've happened. Since we've got a reference on
108 * the bh, even if it commits while we're doing the
109 * copy, the data is still good. */
110 if (buffer_jbd(buffer_cache_bh)
111 && ocfs2_inode_is_new(inode)) {
112 kaddr = kmap_atomic(bh_result->b_page);
114 mlog(ML_ERROR, "couldn't kmap!\n");
117 memcpy(kaddr + (bh_result->b_size * iblock),
118 buffer_cache_bh->b_data,
120 kunmap_atomic(kaddr);
121 set_buffer_uptodate(bh_result);
123 brelse(buffer_cache_bh);
126 map_bh(bh_result, inode->i_sb,
127 le64_to_cpu(fe->id2.i_list.l_recs[0].e_blkno) + iblock);
137 int ocfs2_get_block(struct inode *inode, sector_t iblock,
138 struct buffer_head *bh_result, int create)
141 unsigned int ext_flags;
142 u64 max_blocks = bh_result->b_size >> inode->i_blkbits;
143 u64 p_blkno, count, past_eof;
144 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
146 trace_ocfs2_get_block((unsigned long long)OCFS2_I(inode)->ip_blkno,
147 (unsigned long long)iblock, bh_result, create);
149 if (OCFS2_I(inode)->ip_flags & OCFS2_INODE_SYSTEM_FILE)
150 mlog(ML_NOTICE, "get_block on system inode 0x%p (%lu)\n",
151 inode, inode->i_ino);
153 if (S_ISLNK(inode->i_mode)) {
154 /* this always does I/O for some reason. */
155 err = ocfs2_symlink_get_block(inode, iblock, bh_result, create);
159 err = ocfs2_extent_map_get_blocks(inode, iblock, &p_blkno, &count,
162 mlog(ML_ERROR, "Error %d from get_blocks(0x%p, %llu, 1, "
163 "%llu, NULL)\n", err, inode, (unsigned long long)iblock,
164 (unsigned long long)p_blkno);
168 if (max_blocks < count)
172 * ocfs2 never allocates in this function - the only time we
173 * need to use BH_New is when we're extending i_size on a file
174 * system which doesn't support holes, in which case BH_New
175 * allows __block_write_begin() to zero.
177 * If we see this on a sparse file system, then a truncate has
178 * raced us and removed the cluster. In this case, we clear
179 * the buffers dirty and uptodate bits and let the buffer code
180 * ignore it as a hole.
182 if (create && p_blkno == 0 && ocfs2_sparse_alloc(osb)) {
183 clear_buffer_dirty(bh_result);
184 clear_buffer_uptodate(bh_result);
188 /* Treat the unwritten extent as a hole for zeroing purposes. */
189 if (p_blkno && !(ext_flags & OCFS2_EXT_UNWRITTEN))
190 map_bh(bh_result, inode->i_sb, p_blkno);
192 bh_result->b_size = count << inode->i_blkbits;
194 if (!ocfs2_sparse_alloc(osb)) {
198 "iblock = %llu p_blkno = %llu blkno=(%llu)\n",
199 (unsigned long long)iblock,
200 (unsigned long long)p_blkno,
201 (unsigned long long)OCFS2_I(inode)->ip_blkno);
202 mlog(ML_ERROR, "Size %llu, clusters %u\n", (unsigned long long)i_size_read(inode), OCFS2_I(inode)->ip_clusters);
208 past_eof = ocfs2_blocks_for_bytes(inode->i_sb, i_size_read(inode));
210 trace_ocfs2_get_block_end((unsigned long long)OCFS2_I(inode)->ip_blkno,
211 (unsigned long long)past_eof);
212 if (create && (iblock >= past_eof))
213 set_buffer_new(bh_result);
222 int ocfs2_read_inline_data(struct inode *inode, struct page *page,
223 struct buffer_head *di_bh)
227 struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
229 if (!(le16_to_cpu(di->i_dyn_features) & OCFS2_INLINE_DATA_FL)) {
230 ocfs2_error(inode->i_sb, "Inode %llu lost inline data flag\n",
231 (unsigned long long)OCFS2_I(inode)->ip_blkno);
235 size = i_size_read(inode);
237 if (size > PAGE_CACHE_SIZE ||
238 size > ocfs2_max_inline_data_with_xattr(inode->i_sb, di)) {
239 ocfs2_error(inode->i_sb,
240 "Inode %llu has with inline data has bad size: %Lu\n",
241 (unsigned long long)OCFS2_I(inode)->ip_blkno,
242 (unsigned long long)size);
246 kaddr = kmap_atomic(page);
248 memcpy(kaddr, di->id2.i_data.id_data, size);
249 /* Clear the remaining part of the page */
250 memset(kaddr + size, 0, PAGE_CACHE_SIZE - size);
251 flush_dcache_page(page);
252 kunmap_atomic(kaddr);
254 SetPageUptodate(page);
259 static int ocfs2_readpage_inline(struct inode *inode, struct page *page)
262 struct buffer_head *di_bh = NULL;
264 BUG_ON(!PageLocked(page));
265 BUG_ON(!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL));
267 ret = ocfs2_read_inode_block(inode, &di_bh);
273 ret = ocfs2_read_inline_data(inode, page, di_bh);
281 static int ocfs2_readpage(struct file *file, struct page *page)
283 struct inode *inode = page->mapping->host;
284 struct ocfs2_inode_info *oi = OCFS2_I(inode);
285 loff_t start = (loff_t)page->index << PAGE_CACHE_SHIFT;
288 trace_ocfs2_readpage((unsigned long long)oi->ip_blkno,
289 (page ? page->index : 0));
291 ret = ocfs2_inode_lock_with_page(inode, NULL, 0, page);
293 if (ret == AOP_TRUNCATED_PAGE)
299 if (down_read_trylock(&oi->ip_alloc_sem) == 0) {
301 * Unlock the page and cycle ip_alloc_sem so that we don't
302 * busyloop waiting for ip_alloc_sem to unlock
304 ret = AOP_TRUNCATED_PAGE;
307 down_read(&oi->ip_alloc_sem);
308 up_read(&oi->ip_alloc_sem);
309 goto out_inode_unlock;
313 * i_size might have just been updated as we grabed the meta lock. We
314 * might now be discovering a truncate that hit on another node.
315 * block_read_full_page->get_block freaks out if it is asked to read
316 * beyond the end of a file, so we check here. Callers
317 * (generic_file_read, vm_ops->fault) are clever enough to check i_size
318 * and notice that the page they just read isn't needed.
320 * XXX sys_readahead() seems to get that wrong?
322 if (start >= i_size_read(inode)) {
323 zero_user(page, 0, PAGE_SIZE);
324 SetPageUptodate(page);
329 if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL)
330 ret = ocfs2_readpage_inline(inode, page);
332 ret = block_read_full_page(page, ocfs2_get_block);
336 up_read(&OCFS2_I(inode)->ip_alloc_sem);
338 ocfs2_inode_unlock(inode, 0);
346 * This is used only for read-ahead. Failures or difficult to handle
347 * situations are safe to ignore.
349 * Right now, we don't bother with BH_Boundary - in-inode extent lists
350 * are quite large (243 extents on 4k blocks), so most inodes don't
351 * grow out to a tree. If need be, detecting boundary extents could
352 * trivially be added in a future version of ocfs2_get_block().
354 static int ocfs2_readpages(struct file *filp, struct address_space *mapping,
355 struct list_head *pages, unsigned nr_pages)
358 struct inode *inode = mapping->host;
359 struct ocfs2_inode_info *oi = OCFS2_I(inode);
364 * Use the nonblocking flag for the dlm code to avoid page
365 * lock inversion, but don't bother with retrying.
367 ret = ocfs2_inode_lock_full(inode, NULL, 0, OCFS2_LOCK_NONBLOCK);
371 if (down_read_trylock(&oi->ip_alloc_sem) == 0) {
372 ocfs2_inode_unlock(inode, 0);
377 * Don't bother with inline-data. There isn't anything
378 * to read-ahead in that case anyway...
380 if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL)
384 * Check whether a remote node truncated this file - we just
385 * drop out in that case as it's not worth handling here.
387 last = list_entry(pages->prev, struct page, lru);
388 start = (loff_t)last->index << PAGE_CACHE_SHIFT;
389 if (start >= i_size_read(inode))
392 err = mpage_readpages(mapping, pages, nr_pages, ocfs2_get_block);
395 up_read(&oi->ip_alloc_sem);
396 ocfs2_inode_unlock(inode, 0);
401 /* Note: Because we don't support holes, our allocation has
402 * already happened (allocation writes zeros to the file data)
403 * so we don't have to worry about ordered writes in
406 * ->writepage is called during the process of invalidating the page cache
407 * during blocked lock processing. It can't block on any cluster locks
408 * to during block mapping. It's relying on the fact that the block
409 * mapping can't have disappeared under the dirty pages that it is
410 * being asked to write back.
412 static int ocfs2_writepage(struct page *page, struct writeback_control *wbc)
414 trace_ocfs2_writepage(
415 (unsigned long long)OCFS2_I(page->mapping->host)->ip_blkno,
418 return block_write_full_page(page, ocfs2_get_block, wbc);
421 /* Taken from ext3. We don't necessarily need the full blown
422 * functionality yet, but IMHO it's better to cut and paste the whole
423 * thing so we can avoid introducing our own bugs (and easily pick up
424 * their fixes when they happen) --Mark */
425 int walk_page_buffers( handle_t *handle,
426 struct buffer_head *head,
430 int (*fn)( handle_t *handle,
431 struct buffer_head *bh))
433 struct buffer_head *bh;
434 unsigned block_start, block_end;
435 unsigned blocksize = head->b_size;
437 struct buffer_head *next;
439 for ( bh = head, block_start = 0;
440 ret == 0 && (bh != head || !block_start);
441 block_start = block_end, bh = next)
443 next = bh->b_this_page;
444 block_end = block_start + blocksize;
445 if (block_end <= from || block_start >= to) {
446 if (partial && !buffer_uptodate(bh))
450 err = (*fn)(handle, bh);
457 static sector_t ocfs2_bmap(struct address_space *mapping, sector_t block)
462 struct inode *inode = mapping->host;
464 trace_ocfs2_bmap((unsigned long long)OCFS2_I(inode)->ip_blkno,
465 (unsigned long long)block);
467 /* We don't need to lock journal system files, since they aren't
468 * accessed concurrently from multiple nodes.
470 if (!INODE_JOURNAL(inode)) {
471 err = ocfs2_inode_lock(inode, NULL, 0);
477 down_read(&OCFS2_I(inode)->ip_alloc_sem);
480 if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL))
481 err = ocfs2_extent_map_get_blocks(inode, block, &p_blkno, NULL,
484 if (!INODE_JOURNAL(inode)) {
485 up_read(&OCFS2_I(inode)->ip_alloc_sem);
486 ocfs2_inode_unlock(inode, 0);
490 mlog(ML_ERROR, "get_blocks() failed, block = %llu\n",
491 (unsigned long long)block);
497 status = err ? 0 : p_blkno;
503 * TODO: Make this into a generic get_blocks function.
505 * From do_direct_io in direct-io.c:
506 * "So what we do is to permit the ->get_blocks function to populate
507 * bh.b_size with the size of IO which is permitted at this offset and
510 * This function is called directly from get_more_blocks in direct-io.c.
512 * called like this: dio->get_blocks(dio->inode, fs_startblk,
513 * fs_count, map_bh, dio->rw == WRITE);
515 static int ocfs2_direct_IO_get_blocks(struct inode *inode, sector_t iblock,
516 struct buffer_head *bh_result, int create)
520 int alloc_locked = 0;
521 u64 p_blkno, inode_blocks, contig_blocks;
522 unsigned int ext_flags;
523 unsigned char blocksize_bits = inode->i_sb->s_blocksize_bits;
524 unsigned long max_blocks = bh_result->b_size >> inode->i_blkbits;
525 unsigned long len = bh_result->b_size;
526 unsigned int clusters_to_alloc = 0, contig_clusters = 0;
528 cpos = ocfs2_blocks_to_clusters(inode->i_sb, iblock);
530 /* This function won't even be called if the request isn't all
531 * nicely aligned and of the right size, so there's no need
532 * for us to check any of that. */
534 inode_blocks = ocfs2_blocks_for_bytes(inode->i_sb, i_size_read(inode));
536 down_read(&OCFS2_I(inode)->ip_alloc_sem);
538 /* This figures out the size of the next contiguous block, and
539 * our logical offset */
540 ret = ocfs2_extent_map_get_blocks(inode, iblock, &p_blkno,
541 &contig_blocks, &ext_flags);
542 up_read(&OCFS2_I(inode)->ip_alloc_sem);
545 mlog(ML_ERROR, "get_blocks() failed iblock=%llu\n",
546 (unsigned long long)iblock);
551 /* We should already CoW the refcounted extent in case of create. */
552 BUG_ON(create && (ext_flags & OCFS2_EXT_REFCOUNTED));
554 /* allocate blocks if no p_blkno is found, and create == 1 */
555 if (!p_blkno && create) {
556 ret = ocfs2_inode_lock(inode, NULL, 1);
564 down_write(&OCFS2_I(inode)->ip_alloc_sem);
566 /* fill hole, allocate blocks can't be larger than the size
568 clusters_to_alloc = ocfs2_clusters_for_bytes(inode->i_sb, len);
569 contig_clusters = ocfs2_clusters_for_blocks(inode->i_sb,
571 if (clusters_to_alloc > contig_clusters)
572 clusters_to_alloc = contig_clusters;
574 /* allocate extent and insert them into the extent tree */
575 ret = ocfs2_extend_allocation(inode, cpos,
576 clusters_to_alloc, 0);
578 up_write(&OCFS2_I(inode)->ip_alloc_sem);
583 ret = ocfs2_extent_map_get_blocks(inode, iblock, &p_blkno,
584 &contig_blocks, &ext_flags);
586 up_write(&OCFS2_I(inode)->ip_alloc_sem);
587 mlog(ML_ERROR, "get_blocks() failed iblock=%llu\n",
588 (unsigned long long)iblock);
592 set_buffer_new(bh_result);
593 up_write(&OCFS2_I(inode)->ip_alloc_sem);
597 * get_more_blocks() expects us to describe a hole by clearing
598 * the mapped bit on bh_result().
600 * Consider an unwritten extent as a hole.
602 if (p_blkno && !(ext_flags & OCFS2_EXT_UNWRITTEN))
603 map_bh(bh_result, inode->i_sb, p_blkno);
605 clear_buffer_mapped(bh_result);
607 /* make sure we don't map more than max_blocks blocks here as
608 that's all the kernel will handle at this point. */
609 if (max_blocks < contig_blocks)
610 contig_blocks = max_blocks;
611 bh_result->b_size = contig_blocks << blocksize_bits;
614 ocfs2_inode_unlock(inode, 1);
619 * ocfs2_dio_end_io is called by the dio core when a dio is finished. We're
620 * particularly interested in the aio/dio case. We use the rw_lock DLM lock
621 * to protect io on one node from truncation on another.
623 static int ocfs2_dio_end_io(struct kiocb *iocb,
628 struct inode *inode = file_inode(iocb->ki_filp);
634 /* this io's submitter should not have unlocked this before we could */
635 BUG_ON(!ocfs2_iocb_is_rw_locked(iocb));
637 if (ocfs2_iocb_is_unaligned_aio(iocb)) {
638 ocfs2_iocb_clear_unaligned_aio(iocb);
640 mutex_unlock(&OCFS2_I(inode)->ip_unaligned_aio);
643 /* Let rw unlock to be done later to protect append direct io write */
644 if (offset + bytes <= i_size_read(inode)) {
645 ocfs2_iocb_clear_rw_locked(iocb);
647 level = ocfs2_iocb_rw_locked_level(iocb);
648 ocfs2_rw_unlock(inode, level);
654 static int ocfs2_releasepage(struct page *page, gfp_t wait)
656 if (!page_has_buffers(page))
658 return try_to_free_buffers(page);
661 static int ocfs2_is_overwrite(struct ocfs2_super *osb,
662 struct inode *inode, loff_t offset)
667 unsigned int num_clusters = 0;
668 unsigned int ext_flags = 0;
670 v_cpos = ocfs2_bytes_to_clusters(osb->sb, offset);
671 ret = ocfs2_get_clusters(inode, v_cpos, &p_cpos,
672 &num_clusters, &ext_flags);
678 if (p_cpos && !(ext_flags & OCFS2_EXT_UNWRITTEN))
684 static int ocfs2_direct_IO_zero_extend(struct ocfs2_super *osb,
685 struct inode *inode, loff_t offset,
686 u64 zero_len, int cluster_align)
689 u32 v_cpos = ocfs2_bytes_to_clusters(osb->sb, i_size_read(inode));
690 unsigned int num_clusters = 0;
691 unsigned int ext_flags = 0;
694 if (offset <= i_size_read(inode) || cluster_align)
697 ret = ocfs2_get_clusters(inode, v_cpos, &p_cpos, &num_clusters,
704 if (p_cpos && !(ext_flags & OCFS2_EXT_UNWRITTEN)) {
705 u64 s = i_size_read(inode);
706 sector_t sector = ((u64)p_cpos << (osb->s_clustersize_bits - 9)) +
707 (do_div(s, osb->s_clustersize) >> 9);
709 ret = blkdev_issue_zeroout(osb->sb->s_bdev, sector,
710 zero_len >> 9, GFP_NOFS, false);
718 static int ocfs2_direct_IO_extend_no_holes(struct ocfs2_super *osb,
719 struct inode *inode, loff_t offset)
721 u64 zero_start, zero_len, total_zero_len;
722 u32 p_cpos = 0, clusters_to_add;
723 u32 v_cpos = ocfs2_bytes_to_clusters(osb->sb, i_size_read(inode));
724 unsigned int num_clusters = 0;
725 unsigned int ext_flags = 0;
726 u32 size_div, offset_div;
731 u64 s = i_size_read(inode);
733 offset_div = do_div(o, osb->s_clustersize);
734 size_div = do_div(s, osb->s_clustersize);
737 if (offset <= i_size_read(inode))
740 clusters_to_add = ocfs2_bytes_to_clusters(inode->i_sb, offset) -
741 ocfs2_bytes_to_clusters(inode->i_sb, i_size_read(inode));
742 total_zero_len = offset - i_size_read(inode);
744 total_zero_len -= offset_div;
746 /* Allocate clusters to fill out holes, and this is only needed
747 * when we add more than one clusters. Otherwise the cluster will
748 * be allocated during direct IO */
749 if (clusters_to_add > 1) {
750 ret = ocfs2_extend_allocation(inode,
751 OCFS2_I(inode)->ip_clusters,
752 clusters_to_add - 1, 0);
759 while (total_zero_len) {
760 ret = ocfs2_get_clusters(inode, v_cpos, &p_cpos, &num_clusters,
767 zero_start = ocfs2_clusters_to_bytes(osb->sb, p_cpos) +
769 zero_len = ocfs2_clusters_to_bytes(osb->sb, num_clusters) -
771 zero_len = min(total_zero_len, zero_len);
773 if (p_cpos && !(ext_flags & OCFS2_EXT_UNWRITTEN)) {
774 ret = blkdev_issue_zeroout(osb->sb->s_bdev,
775 zero_start >> 9, zero_len >> 9,
783 total_zero_len -= zero_len;
784 v_cpos += ocfs2_bytes_to_clusters(osb->sb, zero_len + size_div);
786 /* Only at first iteration can be cluster not aligned.
787 * So set size_div to 0 for the rest */
795 static ssize_t ocfs2_direct_IO_write(struct kiocb *iocb,
796 struct iov_iter *iter,
801 bool orphaned = false;
802 int is_overwrite = 0;
803 struct file *file = iocb->ki_filp;
804 struct inode *inode = file_inode(file)->i_mapping->host;
805 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
806 struct buffer_head *di_bh = NULL;
807 size_t count = iter->count;
808 journal_t *journal = osb->journal->j_journal;
809 u64 zero_len_head, zero_len_tail;
810 int cluster_align_head, cluster_align_tail;
811 loff_t final_size = offset + count;
812 int append_write = offset >= i_size_read(inode) ? 1 : 0;
813 unsigned int num_clusters = 0;
814 unsigned int ext_flags = 0;
818 u64 s = i_size_read(inode);
820 zero_len_head = do_div(o, 1 << osb->s_clustersize_bits);
821 cluster_align_head = !zero_len_head;
823 zero_len_tail = osb->s_clustersize -
824 do_div(s, osb->s_clustersize);
825 if ((offset - i_size_read(inode)) < zero_len_tail)
826 zero_len_tail = offset - i_size_read(inode);
827 cluster_align_tail = !zero_len_tail;
831 * when final_size > inode->i_size, inode->i_size will be
832 * updated after direct write, so add the inode to orphan
835 if (final_size > i_size_read(inode)) {
836 ret = ocfs2_add_inode_to_orphan(osb, inode);
845 ret = ocfs2_inode_lock(inode, NULL, 1);
851 /* zeroing out the previously allocated cluster tail
852 * that but not zeroed */
853 if (ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb))) {
854 down_read(&OCFS2_I(inode)->ip_alloc_sem);
855 ret = ocfs2_direct_IO_zero_extend(osb, inode, offset,
856 zero_len_tail, cluster_align_tail);
857 up_read(&OCFS2_I(inode)->ip_alloc_sem);
859 down_write(&OCFS2_I(inode)->ip_alloc_sem);
860 ret = ocfs2_direct_IO_extend_no_holes(osb, inode,
862 up_write(&OCFS2_I(inode)->ip_alloc_sem);
866 ocfs2_inode_unlock(inode, 1);
870 is_overwrite = ocfs2_is_overwrite(osb, inode, offset);
871 if (is_overwrite < 0) {
872 mlog_errno(is_overwrite);
874 ocfs2_inode_unlock(inode, 1);
878 ocfs2_inode_unlock(inode, 1);
881 written = __blockdev_direct_IO(iocb, inode, inode->i_sb->s_bdev, iter,
882 offset, ocfs2_direct_IO_get_blocks,
883 ocfs2_dio_end_io, NULL, 0);
884 /* overwrite aio may return -EIOCBQUEUED, and it is not an error */
885 if ((written < 0) && (written != -EIOCBQUEUED)) {
886 loff_t i_size = i_size_read(inode);
888 if (offset + count > i_size) {
889 ret = ocfs2_inode_lock(inode, &di_bh, 1);
895 if (i_size == i_size_read(inode)) {
896 ret = ocfs2_truncate_file(inode, di_bh,
902 ocfs2_inode_unlock(inode, 1);
909 ocfs2_inode_unlock(inode, 1);
913 ret = jbd2_journal_force_commit(journal);
917 } else if (written > 0 && append_write && !is_overwrite &&
918 !cluster_align_head) {
919 /* zeroing out the allocated cluster head */
921 u32 v_cpos = ocfs2_bytes_to_clusters(osb->sb, offset);
923 ret = ocfs2_inode_lock(inode, NULL, 0);
929 ret = ocfs2_get_clusters(inode, v_cpos, &p_cpos,
930 &num_clusters, &ext_flags);
933 ocfs2_inode_unlock(inode, 0);
937 BUG_ON(!p_cpos || (ext_flags & OCFS2_EXT_UNWRITTEN));
939 ret = blkdev_issue_zeroout(osb->sb->s_bdev,
940 (u64)p_cpos << (osb->s_clustersize_bits - 9),
941 zero_len_head >> 9, GFP_NOFS, false);
945 ocfs2_inode_unlock(inode, 0);
951 int update_isize = written > 0 ? 1 : 0;
952 loff_t end = update_isize ? offset + written : 0;
954 tmp_ret = ocfs2_inode_lock(inode, &di_bh, 1);
961 tmp_ret = ocfs2_del_inode_from_orphan(osb, inode, di_bh,
964 ocfs2_inode_unlock(inode, 1);
971 ocfs2_inode_unlock(inode, 1);
974 tmp_ret = jbd2_journal_force_commit(journal);
987 static ssize_t ocfs2_direct_IO(struct kiocb *iocb, struct iov_iter *iter,
990 struct file *file = iocb->ki_filp;
991 struct inode *inode = file_inode(file)->i_mapping->host;
992 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
993 int full_coherency = !(osb->s_mount_opt &
994 OCFS2_MOUNT_COHERENCY_BUFFERED);
997 * Fallback to buffered I/O if we see an inode without
1000 if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
1003 /* Fallback to buffered I/O if we are appending and
1004 * concurrent O_DIRECT writes are allowed.
1006 if (i_size_read(inode) <= offset && !full_coherency)
1009 if (iov_iter_rw(iter) == READ)
1010 return __blockdev_direct_IO(iocb, inode, inode->i_sb->s_bdev,
1012 ocfs2_direct_IO_get_blocks,
1013 ocfs2_dio_end_io, NULL, 0);
1015 return ocfs2_direct_IO_write(iocb, iter, offset);
1018 static void ocfs2_figure_cluster_boundaries(struct ocfs2_super *osb,
1020 unsigned int *start,
1023 unsigned int cluster_start = 0, cluster_end = PAGE_CACHE_SIZE;
1025 if (unlikely(PAGE_CACHE_SHIFT > osb->s_clustersize_bits)) {
1028 cpp = 1 << (PAGE_CACHE_SHIFT - osb->s_clustersize_bits);
1030 cluster_start = cpos % cpp;
1031 cluster_start = cluster_start << osb->s_clustersize_bits;
1033 cluster_end = cluster_start + osb->s_clustersize;
1036 BUG_ON(cluster_start > PAGE_SIZE);
1037 BUG_ON(cluster_end > PAGE_SIZE);
1040 *start = cluster_start;
1046 * 'from' and 'to' are the region in the page to avoid zeroing.
1048 * If pagesize > clustersize, this function will avoid zeroing outside
1049 * of the cluster boundary.
1051 * from == to == 0 is code for "zero the entire cluster region"
1053 static void ocfs2_clear_page_regions(struct page *page,
1054 struct ocfs2_super *osb, u32 cpos,
1055 unsigned from, unsigned to)
1058 unsigned int cluster_start, cluster_end;
1060 ocfs2_figure_cluster_boundaries(osb, cpos, &cluster_start, &cluster_end);
1062 kaddr = kmap_atomic(page);
1065 if (from > cluster_start)
1066 memset(kaddr + cluster_start, 0, from - cluster_start);
1067 if (to < cluster_end)
1068 memset(kaddr + to, 0, cluster_end - to);
1070 memset(kaddr + cluster_start, 0, cluster_end - cluster_start);
1073 kunmap_atomic(kaddr);
1077 * Nonsparse file systems fully allocate before we get to the write
1078 * code. This prevents ocfs2_write() from tagging the write as an
1079 * allocating one, which means ocfs2_map_page_blocks() might try to
1080 * read-in the blocks at the tail of our file. Avoid reading them by
1081 * testing i_size against each block offset.
1083 static int ocfs2_should_read_blk(struct inode *inode, struct page *page,
1084 unsigned int block_start)
1086 u64 offset = page_offset(page) + block_start;
1088 if (ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)))
1091 if (i_size_read(inode) > offset)
1098 * Some of this taken from __block_write_begin(). We already have our
1099 * mapping by now though, and the entire write will be allocating or
1100 * it won't, so not much need to use BH_New.
1102 * This will also skip zeroing, which is handled externally.
1104 int ocfs2_map_page_blocks(struct page *page, u64 *p_blkno,
1105 struct inode *inode, unsigned int from,
1106 unsigned int to, int new)
1109 struct buffer_head *head, *bh, *wait[2], **wait_bh = wait;
1110 unsigned int block_end, block_start;
1111 unsigned int bsize = 1 << inode->i_blkbits;
1113 if (!page_has_buffers(page))
1114 create_empty_buffers(page, bsize, 0);
1116 head = page_buffers(page);
1117 for (bh = head, block_start = 0; bh != head || !block_start;
1118 bh = bh->b_this_page, block_start += bsize) {
1119 block_end = block_start + bsize;
1121 clear_buffer_new(bh);
1124 * Ignore blocks outside of our i/o range -
1125 * they may belong to unallocated clusters.
1127 if (block_start >= to || block_end <= from) {
1128 if (PageUptodate(page))
1129 set_buffer_uptodate(bh);
1134 * For an allocating write with cluster size >= page
1135 * size, we always write the entire page.
1140 if (!buffer_mapped(bh)) {
1141 map_bh(bh, inode->i_sb, *p_blkno);
1142 unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
1145 if (PageUptodate(page)) {
1146 if (!buffer_uptodate(bh))
1147 set_buffer_uptodate(bh);
1148 } else if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
1150 ocfs2_should_read_blk(inode, page, block_start) &&
1151 (block_start < from || block_end > to)) {
1152 ll_rw_block(READ, 1, &bh);
1156 *p_blkno = *p_blkno + 1;
1160 * If we issued read requests - let them complete.
1162 while(wait_bh > wait) {
1163 wait_on_buffer(*--wait_bh);
1164 if (!buffer_uptodate(*wait_bh))
1168 if (ret == 0 || !new)
1172 * If we get -EIO above, zero out any newly allocated blocks
1173 * to avoid exposing stale data.
1178 block_end = block_start + bsize;
1179 if (block_end <= from)
1181 if (block_start >= to)
1184 zero_user(page, block_start, bh->b_size);
1185 set_buffer_uptodate(bh);
1186 mark_buffer_dirty(bh);
1189 block_start = block_end;
1190 bh = bh->b_this_page;
1191 } while (bh != head);
1196 #if (PAGE_CACHE_SIZE >= OCFS2_MAX_CLUSTERSIZE)
1197 #define OCFS2_MAX_CTXT_PAGES 1
1199 #define OCFS2_MAX_CTXT_PAGES (OCFS2_MAX_CLUSTERSIZE / PAGE_CACHE_SIZE)
1202 #define OCFS2_MAX_CLUSTERS_PER_PAGE (PAGE_CACHE_SIZE / OCFS2_MIN_CLUSTERSIZE)
1205 * Describe the state of a single cluster to be written to.
1207 struct ocfs2_write_cluster_desc {
1211 * Give this a unique field because c_phys eventually gets
1215 unsigned c_unwritten;
1216 unsigned c_needs_zero;
1219 struct ocfs2_write_ctxt {
1220 /* Logical cluster position / len of write */
1224 /* First cluster allocated in a nonsparse extend */
1225 u32 w_first_new_cpos;
1227 /* Type of caller. Must be one of buffer, mmap, direct. */
1228 ocfs2_write_type_t w_type;
1230 struct ocfs2_write_cluster_desc w_desc[OCFS2_MAX_CLUSTERS_PER_PAGE];
1233 * This is true if page_size > cluster_size.
1235 * It triggers a set of special cases during write which might
1236 * have to deal with allocating writes to partial pages.
1238 unsigned int w_large_pages;
1241 * Pages involved in this write.
1243 * w_target_page is the page being written to by the user.
1245 * w_pages is an array of pages which always contains
1246 * w_target_page, and in the case of an allocating write with
1247 * page_size < cluster size, it will contain zero'd and mapped
1248 * pages adjacent to w_target_page which need to be written
1249 * out in so that future reads from that region will get
1252 unsigned int w_num_pages;
1253 struct page *w_pages[OCFS2_MAX_CTXT_PAGES];
1254 struct page *w_target_page;
1257 * w_target_locked is used for page_mkwrite path indicating no unlocking
1258 * against w_target_page in ocfs2_write_end_nolock.
1260 unsigned int w_target_locked:1;
1263 * ocfs2_write_end() uses this to know what the real range to
1264 * write in the target should be.
1266 unsigned int w_target_from;
1267 unsigned int w_target_to;
1270 * We could use journal_current_handle() but this is cleaner,
1275 struct buffer_head *w_di_bh;
1277 struct ocfs2_cached_dealloc_ctxt w_dealloc;
1280 void ocfs2_unlock_and_free_pages(struct page **pages, int num_pages)
1284 for(i = 0; i < num_pages; i++) {
1286 unlock_page(pages[i]);
1287 mark_page_accessed(pages[i]);
1288 page_cache_release(pages[i]);
1293 static void ocfs2_unlock_pages(struct ocfs2_write_ctxt *wc)
1298 * w_target_locked is only set to true in the page_mkwrite() case.
1299 * The intent is to allow us to lock the target page from write_begin()
1300 * to write_end(). The caller must hold a ref on w_target_page.
1302 if (wc->w_target_locked) {
1303 BUG_ON(!wc->w_target_page);
1304 for (i = 0; i < wc->w_num_pages; i++) {
1305 if (wc->w_target_page == wc->w_pages[i]) {
1306 wc->w_pages[i] = NULL;
1310 mark_page_accessed(wc->w_target_page);
1311 page_cache_release(wc->w_target_page);
1313 ocfs2_unlock_and_free_pages(wc->w_pages, wc->w_num_pages);
1316 static void ocfs2_free_write_ctxt(struct ocfs2_write_ctxt *wc)
1318 ocfs2_unlock_pages(wc);
1319 brelse(wc->w_di_bh);
1323 static int ocfs2_alloc_write_ctxt(struct ocfs2_write_ctxt **wcp,
1324 struct ocfs2_super *osb, loff_t pos,
1325 unsigned len, ocfs2_write_type_t type,
1326 struct buffer_head *di_bh)
1329 struct ocfs2_write_ctxt *wc;
1331 wc = kzalloc(sizeof(struct ocfs2_write_ctxt), GFP_NOFS);
1335 wc->w_cpos = pos >> osb->s_clustersize_bits;
1336 wc->w_first_new_cpos = UINT_MAX;
1337 cend = (pos + len - 1) >> osb->s_clustersize_bits;
1338 wc->w_clen = cend - wc->w_cpos + 1;
1340 wc->w_di_bh = di_bh;
1343 if (unlikely(PAGE_CACHE_SHIFT > osb->s_clustersize_bits))
1344 wc->w_large_pages = 1;
1346 wc->w_large_pages = 0;
1348 ocfs2_init_dealloc_ctxt(&wc->w_dealloc);
1356 * If a page has any new buffers, zero them out here, and mark them uptodate
1357 * and dirty so they'll be written out (in order to prevent uninitialised
1358 * block data from leaking). And clear the new bit.
1360 static void ocfs2_zero_new_buffers(struct page *page, unsigned from, unsigned to)
1362 unsigned int block_start, block_end;
1363 struct buffer_head *head, *bh;
1365 BUG_ON(!PageLocked(page));
1366 if (!page_has_buffers(page))
1369 bh = head = page_buffers(page);
1372 block_end = block_start + bh->b_size;
1374 if (buffer_new(bh)) {
1375 if (block_end > from && block_start < to) {
1376 if (!PageUptodate(page)) {
1377 unsigned start, end;
1379 start = max(from, block_start);
1380 end = min(to, block_end);
1382 zero_user_segment(page, start, end);
1383 set_buffer_uptodate(bh);
1386 clear_buffer_new(bh);
1387 mark_buffer_dirty(bh);
1391 block_start = block_end;
1392 bh = bh->b_this_page;
1393 } while (bh != head);
1397 * Only called when we have a failure during allocating write to write
1398 * zero's to the newly allocated region.
1400 static void ocfs2_write_failure(struct inode *inode,
1401 struct ocfs2_write_ctxt *wc,
1402 loff_t user_pos, unsigned user_len)
1405 unsigned from = user_pos & (PAGE_CACHE_SIZE - 1),
1406 to = user_pos + user_len;
1407 struct page *tmppage;
1409 ocfs2_zero_new_buffers(wc->w_target_page, from, to);
1411 for(i = 0; i < wc->w_num_pages; i++) {
1412 tmppage = wc->w_pages[i];
1414 if (page_has_buffers(tmppage)) {
1415 if (ocfs2_should_order_data(inode))
1416 ocfs2_jbd2_file_inode(wc->w_handle, inode);
1418 block_commit_write(tmppage, from, to);
1423 static int ocfs2_prepare_page_for_write(struct inode *inode, u64 *p_blkno,
1424 struct ocfs2_write_ctxt *wc,
1425 struct page *page, u32 cpos,
1426 loff_t user_pos, unsigned user_len,
1430 unsigned int map_from = 0, map_to = 0;
1431 unsigned int cluster_start, cluster_end;
1432 unsigned int user_data_from = 0, user_data_to = 0;
1434 ocfs2_figure_cluster_boundaries(OCFS2_SB(inode->i_sb), cpos,
1435 &cluster_start, &cluster_end);
1437 /* treat the write as new if the a hole/lseek spanned across
1438 * the page boundary.
1440 new = new | ((i_size_read(inode) <= page_offset(page)) &&
1441 (page_offset(page) <= user_pos));
1443 if (page == wc->w_target_page) {
1444 map_from = user_pos & (PAGE_CACHE_SIZE - 1);
1445 map_to = map_from + user_len;
1448 ret = ocfs2_map_page_blocks(page, p_blkno, inode,
1449 cluster_start, cluster_end,
1452 ret = ocfs2_map_page_blocks(page, p_blkno, inode,
1453 map_from, map_to, new);
1459 user_data_from = map_from;
1460 user_data_to = map_to;
1462 map_from = cluster_start;
1463 map_to = cluster_end;
1467 * If we haven't allocated the new page yet, we
1468 * shouldn't be writing it out without copying user
1469 * data. This is likely a math error from the caller.
1473 map_from = cluster_start;
1474 map_to = cluster_end;
1476 ret = ocfs2_map_page_blocks(page, p_blkno, inode,
1477 cluster_start, cluster_end, new);
1485 * Parts of newly allocated pages need to be zero'd.
1487 * Above, we have also rewritten 'to' and 'from' - as far as
1488 * the rest of the function is concerned, the entire cluster
1489 * range inside of a page needs to be written.
1491 * We can skip this if the page is up to date - it's already
1492 * been zero'd from being read in as a hole.
1494 if (new && !PageUptodate(page))
1495 ocfs2_clear_page_regions(page, OCFS2_SB(inode->i_sb),
1496 cpos, user_data_from, user_data_to);
1498 flush_dcache_page(page);
1505 * This function will only grab one clusters worth of pages.
1507 static int ocfs2_grab_pages_for_write(struct address_space *mapping,
1508 struct ocfs2_write_ctxt *wc,
1509 u32 cpos, loff_t user_pos,
1510 unsigned user_len, int new,
1511 struct page *mmap_page)
1514 unsigned long start, target_index, end_index, index;
1515 struct inode *inode = mapping->host;
1518 target_index = user_pos >> PAGE_CACHE_SHIFT;
1521 * Figure out how many pages we'll be manipulating here. For
1522 * non allocating write, we just change the one
1523 * page. Otherwise, we'll need a whole clusters worth. If we're
1524 * writing past i_size, we only need enough pages to cover the
1525 * last page of the write.
1528 wc->w_num_pages = ocfs2_pages_per_cluster(inode->i_sb);
1529 start = ocfs2_align_clusters_to_page_index(inode->i_sb, cpos);
1531 * We need the index *past* the last page we could possibly
1532 * touch. This is the page past the end of the write or
1533 * i_size, whichever is greater.
1535 last_byte = max(user_pos + user_len, i_size_read(inode));
1536 BUG_ON(last_byte < 1);
1537 end_index = ((last_byte - 1) >> PAGE_CACHE_SHIFT) + 1;
1538 if ((start + wc->w_num_pages) > end_index)
1539 wc->w_num_pages = end_index - start;
1541 wc->w_num_pages = 1;
1542 start = target_index;
1545 for(i = 0; i < wc->w_num_pages; i++) {
1548 if (index == target_index && mmap_page) {
1550 * ocfs2_pagemkwrite() is a little different
1551 * and wants us to directly use the page
1554 lock_page(mmap_page);
1556 /* Exit and let the caller retry */
1557 if (mmap_page->mapping != mapping) {
1558 WARN_ON(mmap_page->mapping);
1559 unlock_page(mmap_page);
1564 page_cache_get(mmap_page);
1565 wc->w_pages[i] = mmap_page;
1566 wc->w_target_locked = true;
1568 wc->w_pages[i] = find_or_create_page(mapping, index,
1570 if (!wc->w_pages[i]) {
1576 wait_for_stable_page(wc->w_pages[i]);
1578 if (index == target_index)
1579 wc->w_target_page = wc->w_pages[i];
1583 wc->w_target_locked = false;
1588 * Prepare a single cluster for write one cluster into the file.
1590 static int ocfs2_write_cluster(struct address_space *mapping,
1591 u32 phys, unsigned int unwritten,
1592 unsigned int should_zero,
1593 struct ocfs2_alloc_context *data_ac,
1594 struct ocfs2_alloc_context *meta_ac,
1595 struct ocfs2_write_ctxt *wc, u32 cpos,
1596 loff_t user_pos, unsigned user_len)
1599 u64 v_blkno, p_blkno;
1600 struct inode *inode = mapping->host;
1601 struct ocfs2_extent_tree et;
1603 new = phys == 0 ? 1 : 0;
1608 * This is safe to call with the page locks - it won't take
1609 * any additional semaphores or cluster locks.
1612 ret = ocfs2_add_inode_data(OCFS2_SB(inode->i_sb), inode,
1613 &tmp_pos, 1, 0, wc->w_di_bh,
1614 wc->w_handle, data_ac,
1617 * This shouldn't happen because we must have already
1618 * calculated the correct meta data allocation required. The
1619 * internal tree allocation code should know how to increase
1620 * transaction credits itself.
1622 * If need be, we could handle -EAGAIN for a
1623 * RESTART_TRANS here.
1625 mlog_bug_on_msg(ret == -EAGAIN,
1626 "Inode %llu: EAGAIN return during allocation.\n",
1627 (unsigned long long)OCFS2_I(inode)->ip_blkno);
1632 } else if (unwritten) {
1633 ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode),
1635 ret = ocfs2_mark_extent_written(inode, &et,
1636 wc->w_handle, cpos, 1, phys,
1637 meta_ac, &wc->w_dealloc);
1645 v_blkno = ocfs2_clusters_to_blocks(inode->i_sb, cpos);
1647 v_blkno = user_pos >> inode->i_sb->s_blocksize_bits;
1650 * The only reason this should fail is due to an inability to
1651 * find the extent added.
1653 ret = ocfs2_extent_map_get_blocks(inode, v_blkno, &p_blkno, NULL,
1656 mlog(ML_ERROR, "Get physical blkno failed for inode %llu, "
1657 "at logical block %llu",
1658 (unsigned long long)OCFS2_I(inode)->ip_blkno,
1659 (unsigned long long)v_blkno);
1663 BUG_ON(p_blkno == 0);
1665 for(i = 0; i < wc->w_num_pages; i++) {
1668 tmpret = ocfs2_prepare_page_for_write(inode, &p_blkno, wc,
1669 wc->w_pages[i], cpos,
1680 * We only have cleanup to do in case of allocating write.
1683 ocfs2_write_failure(inode, wc, user_pos, user_len);
1690 static int ocfs2_write_cluster_by_desc(struct address_space *mapping,
1691 struct ocfs2_alloc_context *data_ac,
1692 struct ocfs2_alloc_context *meta_ac,
1693 struct ocfs2_write_ctxt *wc,
1694 loff_t pos, unsigned len)
1698 unsigned int local_len = len;
1699 struct ocfs2_write_cluster_desc *desc;
1700 struct ocfs2_super *osb = OCFS2_SB(mapping->host->i_sb);
1702 for (i = 0; i < wc->w_clen; i++) {
1703 desc = &wc->w_desc[i];
1706 * We have to make sure that the total write passed in
1707 * doesn't extend past a single cluster.
1710 cluster_off = pos & (osb->s_clustersize - 1);
1711 if ((cluster_off + local_len) > osb->s_clustersize)
1712 local_len = osb->s_clustersize - cluster_off;
1714 ret = ocfs2_write_cluster(mapping, desc->c_phys,
1718 wc, desc->c_cpos, pos, local_len);
1734 * ocfs2_write_end() wants to know which parts of the target page it
1735 * should complete the write on. It's easiest to compute them ahead of
1736 * time when a more complete view of the write is available.
1738 static void ocfs2_set_target_boundaries(struct ocfs2_super *osb,
1739 struct ocfs2_write_ctxt *wc,
1740 loff_t pos, unsigned len, int alloc)
1742 struct ocfs2_write_cluster_desc *desc;
1744 wc->w_target_from = pos & (PAGE_CACHE_SIZE - 1);
1745 wc->w_target_to = wc->w_target_from + len;
1751 * Allocating write - we may have different boundaries based
1752 * on page size and cluster size.
1754 * NOTE: We can no longer compute one value from the other as
1755 * the actual write length and user provided length may be
1759 if (wc->w_large_pages) {
1761 * We only care about the 1st and last cluster within
1762 * our range and whether they should be zero'd or not. Either
1763 * value may be extended out to the start/end of a
1764 * newly allocated cluster.
1766 desc = &wc->w_desc[0];
1767 if (desc->c_needs_zero)
1768 ocfs2_figure_cluster_boundaries(osb,
1773 desc = &wc->w_desc[wc->w_clen - 1];
1774 if (desc->c_needs_zero)
1775 ocfs2_figure_cluster_boundaries(osb,
1780 wc->w_target_from = 0;
1781 wc->w_target_to = PAGE_CACHE_SIZE;
1786 * Populate each single-cluster write descriptor in the write context
1787 * with information about the i/o to be done.
1789 * Returns the number of clusters that will have to be allocated, as
1790 * well as a worst case estimate of the number of extent records that
1791 * would have to be created during a write to an unwritten region.
1793 static int ocfs2_populate_write_desc(struct inode *inode,
1794 struct ocfs2_write_ctxt *wc,
1795 unsigned int *clusters_to_alloc,
1796 unsigned int *extents_to_split)
1799 struct ocfs2_write_cluster_desc *desc;
1800 unsigned int num_clusters = 0;
1801 unsigned int ext_flags = 0;
1805 *clusters_to_alloc = 0;
1806 *extents_to_split = 0;
1808 for (i = 0; i < wc->w_clen; i++) {
1809 desc = &wc->w_desc[i];
1810 desc->c_cpos = wc->w_cpos + i;
1812 if (num_clusters == 0) {
1814 * Need to look up the next extent record.
1816 ret = ocfs2_get_clusters(inode, desc->c_cpos, &phys,
1817 &num_clusters, &ext_flags);
1823 /* We should already CoW the refcountd extent. */
1824 BUG_ON(ext_flags & OCFS2_EXT_REFCOUNTED);
1827 * Assume worst case - that we're writing in
1828 * the middle of the extent.
1830 * We can assume that the write proceeds from
1831 * left to right, in which case the extent
1832 * insert code is smart enough to coalesce the
1833 * next splits into the previous records created.
1835 if (ext_flags & OCFS2_EXT_UNWRITTEN)
1836 *extents_to_split = *extents_to_split + 2;
1839 * Only increment phys if it doesn't describe
1846 * If w_first_new_cpos is < UINT_MAX, we have a non-sparse
1847 * file that got extended. w_first_new_cpos tells us
1848 * where the newly allocated clusters are so we can
1851 if (desc->c_cpos >= wc->w_first_new_cpos) {
1853 desc->c_needs_zero = 1;
1856 desc->c_phys = phys;
1859 desc->c_needs_zero = 1;
1860 *clusters_to_alloc = *clusters_to_alloc + 1;
1863 if (ext_flags & OCFS2_EXT_UNWRITTEN) {
1864 desc->c_unwritten = 1;
1865 desc->c_needs_zero = 1;
1876 static int ocfs2_write_begin_inline(struct address_space *mapping,
1877 struct inode *inode,
1878 struct ocfs2_write_ctxt *wc)
1881 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1884 struct ocfs2_dinode *di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1886 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1887 if (IS_ERR(handle)) {
1888 ret = PTR_ERR(handle);
1893 page = find_or_create_page(mapping, 0, GFP_NOFS);
1895 ocfs2_commit_trans(osb, handle);
1901 * If we don't set w_num_pages then this page won't get unlocked
1902 * and freed on cleanup of the write context.
1904 wc->w_pages[0] = wc->w_target_page = page;
1905 wc->w_num_pages = 1;
1907 ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), wc->w_di_bh,
1908 OCFS2_JOURNAL_ACCESS_WRITE);
1910 ocfs2_commit_trans(osb, handle);
1916 if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL))
1917 ocfs2_set_inode_data_inline(inode, di);
1919 if (!PageUptodate(page)) {
1920 ret = ocfs2_read_inline_data(inode, page, wc->w_di_bh);
1922 ocfs2_commit_trans(osb, handle);
1928 wc->w_handle = handle;
1933 int ocfs2_size_fits_inline_data(struct buffer_head *di_bh, u64 new_size)
1935 struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
1937 if (new_size <= le16_to_cpu(di->id2.i_data.id_count))
1942 static int ocfs2_try_to_write_inline_data(struct address_space *mapping,
1943 struct inode *inode, loff_t pos,
1944 unsigned len, struct page *mmap_page,
1945 struct ocfs2_write_ctxt *wc)
1947 int ret, written = 0;
1948 loff_t end = pos + len;
1949 struct ocfs2_inode_info *oi = OCFS2_I(inode);
1950 struct ocfs2_dinode *di = NULL;
1952 trace_ocfs2_try_to_write_inline_data((unsigned long long)oi->ip_blkno,
1953 len, (unsigned long long)pos,
1954 oi->ip_dyn_features);
1957 * Handle inodes which already have inline data 1st.
1959 if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1960 if (mmap_page == NULL &&
1961 ocfs2_size_fits_inline_data(wc->w_di_bh, end))
1962 goto do_inline_write;
1965 * The write won't fit - we have to give this inode an
1966 * inline extent list now.
1968 ret = ocfs2_convert_inline_data_to_extents(inode, wc->w_di_bh);
1975 * Check whether the inode can accept inline data.
1977 if (oi->ip_clusters != 0 || i_size_read(inode) != 0)
1981 * Check whether the write can fit.
1983 di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1985 end > ocfs2_max_inline_data_with_xattr(inode->i_sb, di))
1989 ret = ocfs2_write_begin_inline(mapping, inode, wc);
1996 * This signals to the caller that the data can be written
2001 return written ? written : ret;
2005 * This function only does anything for file systems which can't
2006 * handle sparse files.
2008 * What we want to do here is fill in any hole between the current end
2009 * of allocation and the end of our write. That way the rest of the
2010 * write path can treat it as an non-allocating write, which has no
2011 * special case code for sparse/nonsparse files.
2013 static int ocfs2_expand_nonsparse_inode(struct inode *inode,
2014 struct buffer_head *di_bh,
2015 loff_t pos, unsigned len,
2016 struct ocfs2_write_ctxt *wc)
2019 loff_t newsize = pos + len;
2021 BUG_ON(ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)));
2023 if (newsize <= i_size_read(inode))
2026 ret = ocfs2_extend_no_holes(inode, di_bh, newsize, pos);
2030 wc->w_first_new_cpos =
2031 ocfs2_clusters_for_bytes(inode->i_sb, i_size_read(inode));
2036 static int ocfs2_zero_tail(struct inode *inode, struct buffer_head *di_bh,
2041 BUG_ON(!ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)));
2042 if (pos > i_size_read(inode))
2043 ret = ocfs2_zero_extend(inode, di_bh, pos);
2049 * Try to flush truncate logs if we can free enough clusters from it.
2050 * As for return value, "< 0" means error, "0" no space and "1" means
2051 * we have freed enough spaces and let the caller try to allocate again.
2053 static int ocfs2_try_to_free_truncate_log(struct ocfs2_super *osb,
2054 unsigned int needed)
2058 unsigned int truncated_clusters;
2060 inode_lock(osb->osb_tl_inode);
2061 truncated_clusters = osb->truncated_clusters;
2062 inode_unlock(osb->osb_tl_inode);
2065 * Check whether we can succeed in allocating if we free
2068 if (truncated_clusters < needed)
2071 ret = ocfs2_flush_truncate_log(osb);
2077 if (jbd2_journal_start_commit(osb->journal->j_journal, &target)) {
2078 jbd2_log_wait_commit(osb->journal->j_journal, target);
2085 int ocfs2_write_begin_nolock(struct address_space *mapping,
2086 loff_t pos, unsigned len, ocfs2_write_type_t type,
2087 struct page **pagep, void **fsdata,
2088 struct buffer_head *di_bh, struct page *mmap_page)
2090 int ret, cluster_of_pages, credits = OCFS2_INODE_UPDATE_CREDITS;
2091 unsigned int clusters_to_alloc, extents_to_split, clusters_need = 0;
2092 struct ocfs2_write_ctxt *wc;
2093 struct inode *inode = mapping->host;
2094 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
2095 struct ocfs2_dinode *di;
2096 struct ocfs2_alloc_context *data_ac = NULL;
2097 struct ocfs2_alloc_context *meta_ac = NULL;
2099 struct ocfs2_extent_tree et;
2100 int try_free = 1, ret1;
2103 ret = ocfs2_alloc_write_ctxt(&wc, osb, pos, len, type, di_bh);
2109 if (ocfs2_supports_inline_data(osb)) {
2110 ret = ocfs2_try_to_write_inline_data(mapping, inode, pos, len,
2122 if (ocfs2_sparse_alloc(osb))
2123 ret = ocfs2_zero_tail(inode, di_bh, pos);
2125 ret = ocfs2_expand_nonsparse_inode(inode, di_bh, pos, len,
2132 ret = ocfs2_check_range_for_refcount(inode, pos, len);
2136 } else if (ret == 1) {
2137 clusters_need = wc->w_clen;
2138 ret = ocfs2_refcount_cow(inode, di_bh,
2139 wc->w_cpos, wc->w_clen, UINT_MAX);
2146 ret = ocfs2_populate_write_desc(inode, wc, &clusters_to_alloc,
2152 clusters_need += clusters_to_alloc;
2154 di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
2156 trace_ocfs2_write_begin_nolock(
2157 (unsigned long long)OCFS2_I(inode)->ip_blkno,
2158 (long long)i_size_read(inode),
2159 le32_to_cpu(di->i_clusters),
2160 pos, len, type, mmap_page,
2161 clusters_to_alloc, extents_to_split);
2164 * We set w_target_from, w_target_to here so that
2165 * ocfs2_write_end() knows which range in the target page to
2166 * write out. An allocation requires that we write the entire
2169 if (clusters_to_alloc || extents_to_split) {
2171 * XXX: We are stretching the limits of
2172 * ocfs2_lock_allocators(). It greatly over-estimates
2173 * the work to be done.
2175 ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode),
2177 ret = ocfs2_lock_allocators(inode, &et,
2178 clusters_to_alloc, extents_to_split,
2179 &data_ac, &meta_ac);
2186 data_ac->ac_resv = &OCFS2_I(inode)->ip_la_data_resv;
2188 credits = ocfs2_calc_extend_credits(inode->i_sb,
2194 * We have to zero sparse allocated clusters, unwritten extent clusters,
2195 * and non-sparse clusters we just extended. For non-sparse writes,
2196 * we know zeros will only be needed in the first and/or last cluster.
2198 if (clusters_to_alloc || extents_to_split ||
2199 (wc->w_clen && (wc->w_desc[0].c_needs_zero ||
2200 wc->w_desc[wc->w_clen - 1].c_needs_zero)))
2201 cluster_of_pages = 1;
2203 cluster_of_pages = 0;
2205 ocfs2_set_target_boundaries(osb, wc, pos, len, cluster_of_pages);
2207 handle = ocfs2_start_trans(osb, credits);
2208 if (IS_ERR(handle)) {
2209 ret = PTR_ERR(handle);
2214 wc->w_handle = handle;
2216 if (clusters_to_alloc) {
2217 ret = dquot_alloc_space_nodirty(inode,
2218 ocfs2_clusters_to_bytes(osb->sb, clusters_to_alloc));
2223 ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), wc->w_di_bh,
2224 OCFS2_JOURNAL_ACCESS_WRITE);
2231 * Fill our page array first. That way we've grabbed enough so
2232 * that we can zero and flush if we error after adding the
2235 ret = ocfs2_grab_pages_for_write(mapping, wc, wc->w_cpos, pos, len,
2236 cluster_of_pages, mmap_page);
2237 if (ret && ret != -EAGAIN) {
2243 * ocfs2_grab_pages_for_write() returns -EAGAIN if it could not lock
2244 * the target page. In this case, we exit with no error and no target
2245 * page. This will trigger the caller, page_mkwrite(), to re-try
2248 if (ret == -EAGAIN) {
2249 BUG_ON(wc->w_target_page);
2254 ret = ocfs2_write_cluster_by_desc(mapping, data_ac, meta_ac, wc, pos,
2262 ocfs2_free_alloc_context(data_ac);
2264 ocfs2_free_alloc_context(meta_ac);
2267 *pagep = wc->w_target_page;
2271 if (clusters_to_alloc)
2272 dquot_free_space(inode,
2273 ocfs2_clusters_to_bytes(osb->sb, clusters_to_alloc));
2275 ocfs2_commit_trans(osb, handle);
2278 ocfs2_free_write_ctxt(wc);
2281 ocfs2_free_alloc_context(data_ac);
2285 ocfs2_free_alloc_context(meta_ac);
2289 if (ret == -ENOSPC && try_free) {
2291 * Try to free some truncate log so that we can have enough
2292 * clusters to allocate.
2296 ret1 = ocfs2_try_to_free_truncate_log(osb, clusters_need);
2307 static int ocfs2_write_begin(struct file *file, struct address_space *mapping,
2308 loff_t pos, unsigned len, unsigned flags,
2309 struct page **pagep, void **fsdata)
2312 struct buffer_head *di_bh = NULL;
2313 struct inode *inode = mapping->host;
2315 ret = ocfs2_inode_lock(inode, &di_bh, 1);
2322 * Take alloc sem here to prevent concurrent lookups. That way
2323 * the mapping, zeroing and tree manipulation within
2324 * ocfs2_write() will be safe against ->readpage(). This
2325 * should also serve to lock out allocation from a shared
2328 down_write(&OCFS2_I(inode)->ip_alloc_sem);
2330 ret = ocfs2_write_begin_nolock(mapping, pos, len, OCFS2_WRITE_BUFFER,
2331 pagep, fsdata, di_bh, NULL);
2342 up_write(&OCFS2_I(inode)->ip_alloc_sem);
2345 ocfs2_inode_unlock(inode, 1);
2350 static void ocfs2_write_end_inline(struct inode *inode, loff_t pos,
2351 unsigned len, unsigned *copied,
2352 struct ocfs2_dinode *di,
2353 struct ocfs2_write_ctxt *wc)
2357 if (unlikely(*copied < len)) {
2358 if (!PageUptodate(wc->w_target_page)) {
2364 kaddr = kmap_atomic(wc->w_target_page);
2365 memcpy(di->id2.i_data.id_data + pos, kaddr + pos, *copied);
2366 kunmap_atomic(kaddr);
2368 trace_ocfs2_write_end_inline(
2369 (unsigned long long)OCFS2_I(inode)->ip_blkno,
2370 (unsigned long long)pos, *copied,
2371 le16_to_cpu(di->id2.i_data.id_count),
2372 le16_to_cpu(di->i_dyn_features));
2375 int ocfs2_write_end_nolock(struct address_space *mapping,
2376 loff_t pos, unsigned len, unsigned copied,
2377 struct page *page, void *fsdata)
2380 unsigned from, to, start = pos & (PAGE_CACHE_SIZE - 1);
2381 struct inode *inode = mapping->host;
2382 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
2383 struct ocfs2_write_ctxt *wc = fsdata;
2384 struct ocfs2_dinode *di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
2385 handle_t *handle = wc->w_handle;
2386 struct page *tmppage;
2388 ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), wc->w_di_bh,
2389 OCFS2_JOURNAL_ACCESS_WRITE);
2396 if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
2397 ocfs2_write_end_inline(inode, pos, len, &copied, di, wc);
2398 goto out_write_size;
2401 if (unlikely(copied < len)) {
2402 if (!PageUptodate(wc->w_target_page))
2405 ocfs2_zero_new_buffers(wc->w_target_page, start+copied,
2408 flush_dcache_page(wc->w_target_page);
2410 for(i = 0; i < wc->w_num_pages; i++) {
2411 tmppage = wc->w_pages[i];
2413 if (tmppage == wc->w_target_page) {
2414 from = wc->w_target_from;
2415 to = wc->w_target_to;
2417 BUG_ON(from > PAGE_CACHE_SIZE ||
2418 to > PAGE_CACHE_SIZE ||
2422 * Pages adjacent to the target (if any) imply
2423 * a hole-filling write in which case we want
2424 * to flush their entire range.
2427 to = PAGE_CACHE_SIZE;
2430 if (page_has_buffers(tmppage)) {
2431 if (ocfs2_should_order_data(inode))
2432 ocfs2_jbd2_file_inode(wc->w_handle, inode);
2433 block_commit_write(tmppage, from, to);
2439 if (pos > i_size_read(inode)) {
2440 i_size_write(inode, pos);
2441 mark_inode_dirty(inode);
2443 inode->i_blocks = ocfs2_inode_sector_count(inode);
2444 di->i_size = cpu_to_le64((u64)i_size_read(inode));
2445 inode->i_mtime = inode->i_ctime = CURRENT_TIME;
2446 di->i_mtime = di->i_ctime = cpu_to_le64(inode->i_mtime.tv_sec);
2447 di->i_mtime_nsec = di->i_ctime_nsec = cpu_to_le32(inode->i_mtime.tv_nsec);
2448 ocfs2_update_inode_fsync_trans(handle, inode, 1);
2449 ocfs2_journal_dirty(handle, wc->w_di_bh);
2452 /* unlock pages before dealloc since it needs acquiring j_trans_barrier
2453 * lock, or it will cause a deadlock since journal commit threads holds
2454 * this lock and will ask for the page lock when flushing the data.
2455 * put it here to preserve the unlock order.
2457 ocfs2_unlock_pages(wc);
2459 ocfs2_commit_trans(osb, handle);
2461 ocfs2_run_deallocs(osb, &wc->w_dealloc);
2463 brelse(wc->w_di_bh);
2469 static int ocfs2_write_end(struct file *file, struct address_space *mapping,
2470 loff_t pos, unsigned len, unsigned copied,
2471 struct page *page, void *fsdata)
2474 struct inode *inode = mapping->host;
2476 ret = ocfs2_write_end_nolock(mapping, pos, len, copied, page, fsdata);
2478 up_write(&OCFS2_I(inode)->ip_alloc_sem);
2479 ocfs2_inode_unlock(inode, 1);
2484 const struct address_space_operations ocfs2_aops = {
2485 .readpage = ocfs2_readpage,
2486 .readpages = ocfs2_readpages,
2487 .writepage = ocfs2_writepage,
2488 .write_begin = ocfs2_write_begin,
2489 .write_end = ocfs2_write_end,
2491 .direct_IO = ocfs2_direct_IO,
2492 .invalidatepage = block_invalidatepage,
2493 .releasepage = ocfs2_releasepage,
2494 .migratepage = buffer_migrate_page,
2495 .is_partially_uptodate = block_is_partially_uptodate,
2496 .error_remove_page = generic_error_remove_page,