1 // SPDX-License-Identifier: GPL-2.0-or-later
3 * Copyright (C) 2002, 2004 Oracle. All rights reserved.
7 #include <linux/slab.h>
8 #include <linux/highmem.h>
9 #include <linux/pagemap.h>
10 #include <asm/byteorder.h>
11 #include <linux/swap.h>
12 #include <linux/mpage.h>
13 #include <linux/quotaops.h>
14 #include <linux/blkdev.h>
15 #include <linux/uio.h>
18 #include <cluster/masklog.h>
25 #include "extent_map.h"
32 #include "refcounttree.h"
33 #include "ocfs2_trace.h"
35 #include "buffer_head_io.h"
40 static int ocfs2_symlink_get_block(struct inode *inode, sector_t iblock,
41 struct buffer_head *bh_result, int create)
45 struct ocfs2_dinode *fe = NULL;
46 struct buffer_head *bh = NULL;
47 struct buffer_head *buffer_cache_bh = NULL;
48 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
51 trace_ocfs2_symlink_get_block(
52 (unsigned long long)OCFS2_I(inode)->ip_blkno,
53 (unsigned long long)iblock, bh_result, create);
55 BUG_ON(ocfs2_inode_is_fast_symlink(inode));
57 if ((iblock << inode->i_sb->s_blocksize_bits) > PATH_MAX + 1) {
58 mlog(ML_ERROR, "block offset > PATH_MAX: %llu",
59 (unsigned long long)iblock);
63 status = ocfs2_read_inode_block(inode, &bh);
68 fe = (struct ocfs2_dinode *) bh->b_data;
70 if ((u64)iblock >= ocfs2_clusters_to_blocks(inode->i_sb,
71 le32_to_cpu(fe->i_clusters))) {
73 mlog(ML_ERROR, "block offset is outside the allocated size: "
74 "%llu\n", (unsigned long long)iblock);
78 /* We don't use the page cache to create symlink data, so if
79 * need be, copy it over from the buffer cache. */
80 if (!buffer_uptodate(bh_result) && ocfs2_inode_is_new(inode)) {
81 u64 blkno = le64_to_cpu(fe->id2.i_list.l_recs[0].e_blkno) +
83 buffer_cache_bh = sb_getblk(osb->sb, blkno);
84 if (!buffer_cache_bh) {
86 mlog(ML_ERROR, "couldn't getblock for symlink!\n");
90 /* we haven't locked out transactions, so a commit
91 * could've happened. Since we've got a reference on
92 * the bh, even if it commits while we're doing the
93 * copy, the data is still good. */
94 if (buffer_jbd(buffer_cache_bh)
95 && ocfs2_inode_is_new(inode)) {
96 kaddr = kmap_atomic(bh_result->b_page);
98 mlog(ML_ERROR, "couldn't kmap!\n");
101 memcpy(kaddr + (bh_result->b_size * iblock),
102 buffer_cache_bh->b_data,
104 kunmap_atomic(kaddr);
105 set_buffer_uptodate(bh_result);
107 brelse(buffer_cache_bh);
110 map_bh(bh_result, inode->i_sb,
111 le64_to_cpu(fe->id2.i_list.l_recs[0].e_blkno) + iblock);
121 static int ocfs2_lock_get_block(struct inode *inode, sector_t iblock,
122 struct buffer_head *bh_result, int create)
125 struct ocfs2_inode_info *oi = OCFS2_I(inode);
127 down_read(&oi->ip_alloc_sem);
128 ret = ocfs2_get_block(inode, iblock, bh_result, create);
129 up_read(&oi->ip_alloc_sem);
134 int ocfs2_get_block(struct inode *inode, sector_t iblock,
135 struct buffer_head *bh_result, int create)
138 unsigned int ext_flags;
139 u64 max_blocks = bh_result->b_size >> inode->i_blkbits;
140 u64 p_blkno, count, past_eof;
141 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
143 trace_ocfs2_get_block((unsigned long long)OCFS2_I(inode)->ip_blkno,
144 (unsigned long long)iblock, bh_result, create);
146 if (OCFS2_I(inode)->ip_flags & OCFS2_INODE_SYSTEM_FILE)
147 mlog(ML_NOTICE, "get_block on system inode 0x%p (%lu)\n",
148 inode, inode->i_ino);
150 if (S_ISLNK(inode->i_mode)) {
151 /* this always does I/O for some reason. */
152 err = ocfs2_symlink_get_block(inode, iblock, bh_result, create);
156 err = ocfs2_extent_map_get_blocks(inode, iblock, &p_blkno, &count,
159 mlog(ML_ERROR, "Error %d from get_blocks(0x%p, %llu, 1, "
160 "%llu, NULL)\n", err, inode, (unsigned long long)iblock,
161 (unsigned long long)p_blkno);
165 if (max_blocks < count)
169 * ocfs2 never allocates in this function - the only time we
170 * need to use BH_New is when we're extending i_size on a file
171 * system which doesn't support holes, in which case BH_New
172 * allows __block_write_begin() to zero.
174 * If we see this on a sparse file system, then a truncate has
175 * raced us and removed the cluster. In this case, we clear
176 * the buffers dirty and uptodate bits and let the buffer code
177 * ignore it as a hole.
179 if (create && p_blkno == 0 && ocfs2_sparse_alloc(osb)) {
180 clear_buffer_dirty(bh_result);
181 clear_buffer_uptodate(bh_result);
185 /* Treat the unwritten extent as a hole for zeroing purposes. */
186 if (p_blkno && !(ext_flags & OCFS2_EXT_UNWRITTEN))
187 map_bh(bh_result, inode->i_sb, p_blkno);
189 bh_result->b_size = count << inode->i_blkbits;
191 if (!ocfs2_sparse_alloc(osb)) {
195 "iblock = %llu p_blkno = %llu blkno=(%llu)\n",
196 (unsigned long long)iblock,
197 (unsigned long long)p_blkno,
198 (unsigned long long)OCFS2_I(inode)->ip_blkno);
199 mlog(ML_ERROR, "Size %llu, clusters %u\n", (unsigned long long)i_size_read(inode), OCFS2_I(inode)->ip_clusters);
205 past_eof = ocfs2_blocks_for_bytes(inode->i_sb, i_size_read(inode));
207 trace_ocfs2_get_block_end((unsigned long long)OCFS2_I(inode)->ip_blkno,
208 (unsigned long long)past_eof);
209 if (create && (iblock >= past_eof))
210 set_buffer_new(bh_result);
219 int ocfs2_read_inline_data(struct inode *inode, struct page *page,
220 struct buffer_head *di_bh)
224 struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
226 if (!(le16_to_cpu(di->i_dyn_features) & OCFS2_INLINE_DATA_FL)) {
227 ocfs2_error(inode->i_sb, "Inode %llu lost inline data flag\n",
228 (unsigned long long)OCFS2_I(inode)->ip_blkno);
232 size = i_size_read(inode);
234 if (size > PAGE_SIZE ||
235 size > ocfs2_max_inline_data_with_xattr(inode->i_sb, di)) {
236 ocfs2_error(inode->i_sb,
237 "Inode %llu has with inline data has bad size: %Lu\n",
238 (unsigned long long)OCFS2_I(inode)->ip_blkno,
239 (unsigned long long)size);
243 kaddr = kmap_atomic(page);
245 memcpy(kaddr, di->id2.i_data.id_data, size);
246 /* Clear the remaining part of the page */
247 memset(kaddr + size, 0, PAGE_SIZE - size);
248 flush_dcache_page(page);
249 kunmap_atomic(kaddr);
251 SetPageUptodate(page);
256 static int ocfs2_readpage_inline(struct inode *inode, struct page *page)
259 struct buffer_head *di_bh = NULL;
261 BUG_ON(!PageLocked(page));
262 BUG_ON(!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL));
264 ret = ocfs2_read_inode_block(inode, &di_bh);
270 ret = ocfs2_read_inline_data(inode, page, di_bh);
278 static int ocfs2_read_folio(struct file *file, struct folio *folio)
280 struct page *page = &folio->page;
281 struct inode *inode = page->mapping->host;
282 struct ocfs2_inode_info *oi = OCFS2_I(inode);
283 loff_t start = (loff_t)page->index << PAGE_SHIFT;
286 trace_ocfs2_readpage((unsigned long long)oi->ip_blkno,
287 (page ? page->index : 0));
289 ret = ocfs2_inode_lock_with_page(inode, NULL, 0, page);
291 if (ret == AOP_TRUNCATED_PAGE)
297 if (down_read_trylock(&oi->ip_alloc_sem) == 0) {
299 * Unlock the page and cycle ip_alloc_sem so that we don't
300 * busyloop waiting for ip_alloc_sem to unlock
302 ret = AOP_TRUNCATED_PAGE;
305 down_read(&oi->ip_alloc_sem);
306 up_read(&oi->ip_alloc_sem);
307 goto out_inode_unlock;
311 * i_size might have just been updated as we grabed the meta lock. We
312 * might now be discovering a truncate that hit on another node.
313 * block_read_full_folio->get_block freaks out if it is asked to read
314 * beyond the end of a file, so we check here. Callers
315 * (generic_file_read, vm_ops->fault) are clever enough to check i_size
316 * and notice that the page they just read isn't needed.
318 * XXX sys_readahead() seems to get that wrong?
320 if (start >= i_size_read(inode)) {
321 zero_user(page, 0, PAGE_SIZE);
322 SetPageUptodate(page);
327 if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL)
328 ret = ocfs2_readpage_inline(inode, page);
330 ret = block_read_full_folio(page_folio(page), ocfs2_get_block);
334 up_read(&oi->ip_alloc_sem);
336 ocfs2_inode_unlock(inode, 0);
344 * This is used only for read-ahead. Failures or difficult to handle
345 * situations are safe to ignore.
347 * Right now, we don't bother with BH_Boundary - in-inode extent lists
348 * are quite large (243 extents on 4k blocks), so most inodes don't
349 * grow out to a tree. If need be, detecting boundary extents could
350 * trivially be added in a future version of ocfs2_get_block().
352 static void ocfs2_readahead(struct readahead_control *rac)
355 struct inode *inode = rac->mapping->host;
356 struct ocfs2_inode_info *oi = OCFS2_I(inode);
359 * Use the nonblocking flag for the dlm code to avoid page
360 * lock inversion, but don't bother with retrying.
362 ret = ocfs2_inode_lock_full(inode, NULL, 0, OCFS2_LOCK_NONBLOCK);
366 if (down_read_trylock(&oi->ip_alloc_sem) == 0)
370 * Don't bother with inline-data. There isn't anything
371 * to read-ahead in that case anyway...
373 if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL)
377 * Check whether a remote node truncated this file - we just
378 * drop out in that case as it's not worth handling here.
380 if (readahead_pos(rac) >= i_size_read(inode))
383 mpage_readahead(rac, ocfs2_get_block);
386 up_read(&oi->ip_alloc_sem);
388 ocfs2_inode_unlock(inode, 0);
391 /* Note: Because we don't support holes, our allocation has
392 * already happened (allocation writes zeros to the file data)
393 * so we don't have to worry about ordered writes in
396 * ->writepage is called during the process of invalidating the page cache
397 * during blocked lock processing. It can't block on any cluster locks
398 * to during block mapping. It's relying on the fact that the block
399 * mapping can't have disappeared under the dirty pages that it is
400 * being asked to write back.
402 static int ocfs2_writepage(struct page *page, struct writeback_control *wbc)
404 trace_ocfs2_writepage(
405 (unsigned long long)OCFS2_I(page->mapping->host)->ip_blkno,
408 return block_write_full_page(page, ocfs2_get_block, wbc);
411 /* Taken from ext3. We don't necessarily need the full blown
412 * functionality yet, but IMHO it's better to cut and paste the whole
413 * thing so we can avoid introducing our own bugs (and easily pick up
414 * their fixes when they happen) --Mark */
415 int walk_page_buffers( handle_t *handle,
416 struct buffer_head *head,
420 int (*fn)( handle_t *handle,
421 struct buffer_head *bh))
423 struct buffer_head *bh;
424 unsigned block_start, block_end;
425 unsigned blocksize = head->b_size;
427 struct buffer_head *next;
429 for ( bh = head, block_start = 0;
430 ret == 0 && (bh != head || !block_start);
431 block_start = block_end, bh = next)
433 next = bh->b_this_page;
434 block_end = block_start + blocksize;
435 if (block_end <= from || block_start >= to) {
436 if (partial && !buffer_uptodate(bh))
440 err = (*fn)(handle, bh);
447 static sector_t ocfs2_bmap(struct address_space *mapping, sector_t block)
452 struct inode *inode = mapping->host;
454 trace_ocfs2_bmap((unsigned long long)OCFS2_I(inode)->ip_blkno,
455 (unsigned long long)block);
458 * The swap code (ab-)uses ->bmap to get a block mapping and then
459 * bypasseѕ the file system for actual I/O. We really can't allow
460 * that on refcounted inodes, so we have to skip out here. And yes,
461 * 0 is the magic code for a bmap error..
463 if (ocfs2_is_refcount_inode(inode))
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;
501 static bool ocfs2_release_folio(struct folio *folio, gfp_t wait)
503 if (!folio_buffers(folio))
505 return try_to_free_buffers(folio);
508 static void ocfs2_figure_cluster_boundaries(struct ocfs2_super *osb,
513 unsigned int cluster_start = 0, cluster_end = PAGE_SIZE;
515 if (unlikely(PAGE_SHIFT > osb->s_clustersize_bits)) {
518 cpp = 1 << (PAGE_SHIFT - osb->s_clustersize_bits);
520 cluster_start = cpos % cpp;
521 cluster_start = cluster_start << osb->s_clustersize_bits;
523 cluster_end = cluster_start + osb->s_clustersize;
526 BUG_ON(cluster_start > PAGE_SIZE);
527 BUG_ON(cluster_end > PAGE_SIZE);
530 *start = cluster_start;
536 * 'from' and 'to' are the region in the page to avoid zeroing.
538 * If pagesize > clustersize, this function will avoid zeroing outside
539 * of the cluster boundary.
541 * from == to == 0 is code for "zero the entire cluster region"
543 static void ocfs2_clear_page_regions(struct page *page,
544 struct ocfs2_super *osb, u32 cpos,
545 unsigned from, unsigned to)
548 unsigned int cluster_start, cluster_end;
550 ocfs2_figure_cluster_boundaries(osb, cpos, &cluster_start, &cluster_end);
552 kaddr = kmap_atomic(page);
555 if (from > cluster_start)
556 memset(kaddr + cluster_start, 0, from - cluster_start);
557 if (to < cluster_end)
558 memset(kaddr + to, 0, cluster_end - to);
560 memset(kaddr + cluster_start, 0, cluster_end - cluster_start);
563 kunmap_atomic(kaddr);
567 * Nonsparse file systems fully allocate before we get to the write
568 * code. This prevents ocfs2_write() from tagging the write as an
569 * allocating one, which means ocfs2_map_page_blocks() might try to
570 * read-in the blocks at the tail of our file. Avoid reading them by
571 * testing i_size against each block offset.
573 static int ocfs2_should_read_blk(struct inode *inode, struct page *page,
574 unsigned int block_start)
576 u64 offset = page_offset(page) + block_start;
578 if (ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)))
581 if (i_size_read(inode) > offset)
588 * Some of this taken from __block_write_begin(). We already have our
589 * mapping by now though, and the entire write will be allocating or
590 * it won't, so not much need to use BH_New.
592 * This will also skip zeroing, which is handled externally.
594 int ocfs2_map_page_blocks(struct page *page, u64 *p_blkno,
595 struct inode *inode, unsigned int from,
596 unsigned int to, int new)
599 struct buffer_head *head, *bh, *wait[2], **wait_bh = wait;
600 unsigned int block_end, block_start;
601 unsigned int bsize = i_blocksize(inode);
603 if (!page_has_buffers(page))
604 create_empty_buffers(page, bsize, 0);
606 head = page_buffers(page);
607 for (bh = head, block_start = 0; bh != head || !block_start;
608 bh = bh->b_this_page, block_start += bsize) {
609 block_end = block_start + bsize;
611 clear_buffer_new(bh);
614 * Ignore blocks outside of our i/o range -
615 * they may belong to unallocated clusters.
617 if (block_start >= to || block_end <= from) {
618 if (PageUptodate(page))
619 set_buffer_uptodate(bh);
624 * For an allocating write with cluster size >= page
625 * size, we always write the entire page.
630 if (!buffer_mapped(bh)) {
631 map_bh(bh, inode->i_sb, *p_blkno);
632 clean_bdev_bh_alias(bh);
635 if (PageUptodate(page)) {
636 set_buffer_uptodate(bh);
637 } else if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
639 ocfs2_should_read_blk(inode, page, block_start) &&
640 (block_start < from || block_end > to)) {
641 ll_rw_block(REQ_OP_READ, 1, &bh);
645 *p_blkno = *p_blkno + 1;
649 * If we issued read requests - let them complete.
651 while(wait_bh > wait) {
652 wait_on_buffer(*--wait_bh);
653 if (!buffer_uptodate(*wait_bh))
657 if (ret == 0 || !new)
661 * If we get -EIO above, zero out any newly allocated blocks
662 * to avoid exposing stale data.
667 block_end = block_start + bsize;
668 if (block_end <= from)
670 if (block_start >= to)
673 zero_user(page, block_start, bh->b_size);
674 set_buffer_uptodate(bh);
675 mark_buffer_dirty(bh);
678 block_start = block_end;
679 bh = bh->b_this_page;
680 } while (bh != head);
685 #if (PAGE_SIZE >= OCFS2_MAX_CLUSTERSIZE)
686 #define OCFS2_MAX_CTXT_PAGES 1
688 #define OCFS2_MAX_CTXT_PAGES (OCFS2_MAX_CLUSTERSIZE / PAGE_SIZE)
691 #define OCFS2_MAX_CLUSTERS_PER_PAGE (PAGE_SIZE / OCFS2_MIN_CLUSTERSIZE)
693 struct ocfs2_unwritten_extent {
694 struct list_head ue_node;
695 struct list_head ue_ip_node;
701 * Describe the state of a single cluster to be written to.
703 struct ocfs2_write_cluster_desc {
707 * Give this a unique field because c_phys eventually gets
711 unsigned c_clear_unwritten;
712 unsigned c_needs_zero;
715 struct ocfs2_write_ctxt {
716 /* Logical cluster position / len of write */
720 /* First cluster allocated in a nonsparse extend */
721 u32 w_first_new_cpos;
723 /* Type of caller. Must be one of buffer, mmap, direct. */
724 ocfs2_write_type_t w_type;
726 struct ocfs2_write_cluster_desc w_desc[OCFS2_MAX_CLUSTERS_PER_PAGE];
729 * This is true if page_size > cluster_size.
731 * It triggers a set of special cases during write which might
732 * have to deal with allocating writes to partial pages.
734 unsigned int w_large_pages;
737 * Pages involved in this write.
739 * w_target_page is the page being written to by the user.
741 * w_pages is an array of pages which always contains
742 * w_target_page, and in the case of an allocating write with
743 * page_size < cluster size, it will contain zero'd and mapped
744 * pages adjacent to w_target_page which need to be written
745 * out in so that future reads from that region will get
748 unsigned int w_num_pages;
749 struct page *w_pages[OCFS2_MAX_CTXT_PAGES];
750 struct page *w_target_page;
753 * w_target_locked is used for page_mkwrite path indicating no unlocking
754 * against w_target_page in ocfs2_write_end_nolock.
756 unsigned int w_target_locked:1;
759 * ocfs2_write_end() uses this to know what the real range to
760 * write in the target should be.
762 unsigned int w_target_from;
763 unsigned int w_target_to;
766 * We could use journal_current_handle() but this is cleaner,
771 struct buffer_head *w_di_bh;
773 struct ocfs2_cached_dealloc_ctxt w_dealloc;
775 struct list_head w_unwritten_list;
776 unsigned int w_unwritten_count;
779 void ocfs2_unlock_and_free_pages(struct page **pages, int num_pages)
783 for(i = 0; i < num_pages; i++) {
785 unlock_page(pages[i]);
786 mark_page_accessed(pages[i]);
792 static void ocfs2_unlock_pages(struct ocfs2_write_ctxt *wc)
797 * w_target_locked is only set to true in the page_mkwrite() case.
798 * The intent is to allow us to lock the target page from write_begin()
799 * to write_end(). The caller must hold a ref on w_target_page.
801 if (wc->w_target_locked) {
802 BUG_ON(!wc->w_target_page);
803 for (i = 0; i < wc->w_num_pages; i++) {
804 if (wc->w_target_page == wc->w_pages[i]) {
805 wc->w_pages[i] = NULL;
809 mark_page_accessed(wc->w_target_page);
810 put_page(wc->w_target_page);
812 ocfs2_unlock_and_free_pages(wc->w_pages, wc->w_num_pages);
815 static void ocfs2_free_unwritten_list(struct inode *inode,
816 struct list_head *head)
818 struct ocfs2_inode_info *oi = OCFS2_I(inode);
819 struct ocfs2_unwritten_extent *ue = NULL, *tmp = NULL;
821 list_for_each_entry_safe(ue, tmp, head, ue_node) {
822 list_del(&ue->ue_node);
823 spin_lock(&oi->ip_lock);
824 list_del(&ue->ue_ip_node);
825 spin_unlock(&oi->ip_lock);
830 static void ocfs2_free_write_ctxt(struct inode *inode,
831 struct ocfs2_write_ctxt *wc)
833 ocfs2_free_unwritten_list(inode, &wc->w_unwritten_list);
834 ocfs2_unlock_pages(wc);
839 static int ocfs2_alloc_write_ctxt(struct ocfs2_write_ctxt **wcp,
840 struct ocfs2_super *osb, loff_t pos,
841 unsigned len, ocfs2_write_type_t type,
842 struct buffer_head *di_bh)
845 struct ocfs2_write_ctxt *wc;
847 wc = kzalloc(sizeof(struct ocfs2_write_ctxt), GFP_NOFS);
851 wc->w_cpos = pos >> osb->s_clustersize_bits;
852 wc->w_first_new_cpos = UINT_MAX;
853 cend = (pos + len - 1) >> osb->s_clustersize_bits;
854 wc->w_clen = cend - wc->w_cpos + 1;
859 if (unlikely(PAGE_SHIFT > osb->s_clustersize_bits))
860 wc->w_large_pages = 1;
862 wc->w_large_pages = 0;
864 ocfs2_init_dealloc_ctxt(&wc->w_dealloc);
865 INIT_LIST_HEAD(&wc->w_unwritten_list);
873 * If a page has any new buffers, zero them out here, and mark them uptodate
874 * and dirty so they'll be written out (in order to prevent uninitialised
875 * block data from leaking). And clear the new bit.
877 static void ocfs2_zero_new_buffers(struct page *page, unsigned from, unsigned to)
879 unsigned int block_start, block_end;
880 struct buffer_head *head, *bh;
882 BUG_ON(!PageLocked(page));
883 if (!page_has_buffers(page))
886 bh = head = page_buffers(page);
889 block_end = block_start + bh->b_size;
891 if (buffer_new(bh)) {
892 if (block_end > from && block_start < to) {
893 if (!PageUptodate(page)) {
896 start = max(from, block_start);
897 end = min(to, block_end);
899 zero_user_segment(page, start, end);
900 set_buffer_uptodate(bh);
903 clear_buffer_new(bh);
904 mark_buffer_dirty(bh);
908 block_start = block_end;
909 bh = bh->b_this_page;
910 } while (bh != head);
914 * Only called when we have a failure during allocating write to write
915 * zero's to the newly allocated region.
917 static void ocfs2_write_failure(struct inode *inode,
918 struct ocfs2_write_ctxt *wc,
919 loff_t user_pos, unsigned user_len)
922 unsigned from = user_pos & (PAGE_SIZE - 1),
923 to = user_pos + user_len;
924 struct page *tmppage;
926 if (wc->w_target_page)
927 ocfs2_zero_new_buffers(wc->w_target_page, from, to);
929 for(i = 0; i < wc->w_num_pages; i++) {
930 tmppage = wc->w_pages[i];
932 if (tmppage && page_has_buffers(tmppage)) {
933 if (ocfs2_should_order_data(inode))
934 ocfs2_jbd2_inode_add_write(wc->w_handle, inode,
937 block_commit_write(tmppage, from, to);
942 static int ocfs2_prepare_page_for_write(struct inode *inode, u64 *p_blkno,
943 struct ocfs2_write_ctxt *wc,
944 struct page *page, u32 cpos,
945 loff_t user_pos, unsigned user_len,
949 unsigned int map_from = 0, map_to = 0;
950 unsigned int cluster_start, cluster_end;
951 unsigned int user_data_from = 0, user_data_to = 0;
953 ocfs2_figure_cluster_boundaries(OCFS2_SB(inode->i_sb), cpos,
954 &cluster_start, &cluster_end);
956 /* treat the write as new if the a hole/lseek spanned across
959 new = new | ((i_size_read(inode) <= page_offset(page)) &&
960 (page_offset(page) <= user_pos));
962 if (page == wc->w_target_page) {
963 map_from = user_pos & (PAGE_SIZE - 1);
964 map_to = map_from + user_len;
967 ret = ocfs2_map_page_blocks(page, p_blkno, inode,
968 cluster_start, cluster_end,
971 ret = ocfs2_map_page_blocks(page, p_blkno, inode,
972 map_from, map_to, new);
978 user_data_from = map_from;
979 user_data_to = map_to;
981 map_from = cluster_start;
982 map_to = cluster_end;
986 * If we haven't allocated the new page yet, we
987 * shouldn't be writing it out without copying user
988 * data. This is likely a math error from the caller.
992 map_from = cluster_start;
993 map_to = cluster_end;
995 ret = ocfs2_map_page_blocks(page, p_blkno, inode,
996 cluster_start, cluster_end, new);
1004 * Parts of newly allocated pages need to be zero'd.
1006 * Above, we have also rewritten 'to' and 'from' - as far as
1007 * the rest of the function is concerned, the entire cluster
1008 * range inside of a page needs to be written.
1010 * We can skip this if the page is up to date - it's already
1011 * been zero'd from being read in as a hole.
1013 if (new && !PageUptodate(page))
1014 ocfs2_clear_page_regions(page, OCFS2_SB(inode->i_sb),
1015 cpos, user_data_from, user_data_to);
1017 flush_dcache_page(page);
1024 * This function will only grab one clusters worth of pages.
1026 static int ocfs2_grab_pages_for_write(struct address_space *mapping,
1027 struct ocfs2_write_ctxt *wc,
1028 u32 cpos, loff_t user_pos,
1029 unsigned user_len, int new,
1030 struct page *mmap_page)
1033 unsigned long start, target_index, end_index, index;
1034 struct inode *inode = mapping->host;
1037 target_index = user_pos >> PAGE_SHIFT;
1040 * Figure out how many pages we'll be manipulating here. For
1041 * non allocating write, we just change the one
1042 * page. Otherwise, we'll need a whole clusters worth. If we're
1043 * writing past i_size, we only need enough pages to cover the
1044 * last page of the write.
1047 wc->w_num_pages = ocfs2_pages_per_cluster(inode->i_sb);
1048 start = ocfs2_align_clusters_to_page_index(inode->i_sb, cpos);
1050 * We need the index *past* the last page we could possibly
1051 * touch. This is the page past the end of the write or
1052 * i_size, whichever is greater.
1054 last_byte = max(user_pos + user_len, i_size_read(inode));
1055 BUG_ON(last_byte < 1);
1056 end_index = ((last_byte - 1) >> PAGE_SHIFT) + 1;
1057 if ((start + wc->w_num_pages) > end_index)
1058 wc->w_num_pages = end_index - start;
1060 wc->w_num_pages = 1;
1061 start = target_index;
1063 end_index = (user_pos + user_len - 1) >> PAGE_SHIFT;
1065 for(i = 0; i < wc->w_num_pages; i++) {
1068 if (index >= target_index && index <= end_index &&
1069 wc->w_type == OCFS2_WRITE_MMAP) {
1071 * ocfs2_pagemkwrite() is a little different
1072 * and wants us to directly use the page
1075 lock_page(mmap_page);
1077 /* Exit and let the caller retry */
1078 if (mmap_page->mapping != mapping) {
1079 WARN_ON(mmap_page->mapping);
1080 unlock_page(mmap_page);
1085 get_page(mmap_page);
1086 wc->w_pages[i] = mmap_page;
1087 wc->w_target_locked = true;
1088 } else if (index >= target_index && index <= end_index &&
1089 wc->w_type == OCFS2_WRITE_DIRECT) {
1090 /* Direct write has no mapping page. */
1091 wc->w_pages[i] = NULL;
1094 wc->w_pages[i] = find_or_create_page(mapping, index,
1096 if (!wc->w_pages[i]) {
1102 wait_for_stable_page(wc->w_pages[i]);
1104 if (index == target_index)
1105 wc->w_target_page = wc->w_pages[i];
1109 wc->w_target_locked = false;
1114 * Prepare a single cluster for write one cluster into the file.
1116 static int ocfs2_write_cluster(struct address_space *mapping,
1117 u32 *phys, unsigned int new,
1118 unsigned int clear_unwritten,
1119 unsigned int should_zero,
1120 struct ocfs2_alloc_context *data_ac,
1121 struct ocfs2_alloc_context *meta_ac,
1122 struct ocfs2_write_ctxt *wc, u32 cpos,
1123 loff_t user_pos, unsigned user_len)
1127 struct inode *inode = mapping->host;
1128 struct ocfs2_extent_tree et;
1129 int bpc = ocfs2_clusters_to_blocks(inode->i_sb, 1);
1135 * This is safe to call with the page locks - it won't take
1136 * any additional semaphores or cluster locks.
1139 ret = ocfs2_add_inode_data(OCFS2_SB(inode->i_sb), inode,
1140 &tmp_pos, 1, !clear_unwritten,
1141 wc->w_di_bh, wc->w_handle,
1142 data_ac, meta_ac, NULL);
1144 * This shouldn't happen because we must have already
1145 * calculated the correct meta data allocation required. The
1146 * internal tree allocation code should know how to increase
1147 * transaction credits itself.
1149 * If need be, we could handle -EAGAIN for a
1150 * RESTART_TRANS here.
1152 mlog_bug_on_msg(ret == -EAGAIN,
1153 "Inode %llu: EAGAIN return during allocation.\n",
1154 (unsigned long long)OCFS2_I(inode)->ip_blkno);
1159 } else if (clear_unwritten) {
1160 ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode),
1162 ret = ocfs2_mark_extent_written(inode, &et,
1163 wc->w_handle, cpos, 1, *phys,
1164 meta_ac, &wc->w_dealloc);
1172 * The only reason this should fail is due to an inability to
1173 * find the extent added.
1175 ret = ocfs2_get_clusters(inode, cpos, phys, NULL, NULL);
1177 mlog(ML_ERROR, "Get physical blkno failed for inode %llu, "
1178 "at logical cluster %u",
1179 (unsigned long long)OCFS2_I(inode)->ip_blkno, cpos);
1185 p_blkno = ocfs2_clusters_to_blocks(inode->i_sb, *phys);
1187 p_blkno += (user_pos >> inode->i_sb->s_blocksize_bits) & (u64)(bpc - 1);
1189 for(i = 0; i < wc->w_num_pages; i++) {
1192 /* This is the direct io target page. */
1193 if (wc->w_pages[i] == NULL) {
1198 tmpret = ocfs2_prepare_page_for_write(inode, &p_blkno, wc,
1199 wc->w_pages[i], cpos,
1210 * We only have cleanup to do in case of allocating write.
1213 ocfs2_write_failure(inode, wc, user_pos, user_len);
1220 static int ocfs2_write_cluster_by_desc(struct address_space *mapping,
1221 struct ocfs2_alloc_context *data_ac,
1222 struct ocfs2_alloc_context *meta_ac,
1223 struct ocfs2_write_ctxt *wc,
1224 loff_t pos, unsigned len)
1228 unsigned int local_len = len;
1229 struct ocfs2_write_cluster_desc *desc;
1230 struct ocfs2_super *osb = OCFS2_SB(mapping->host->i_sb);
1232 for (i = 0; i < wc->w_clen; i++) {
1233 desc = &wc->w_desc[i];
1236 * We have to make sure that the total write passed in
1237 * doesn't extend past a single cluster.
1240 cluster_off = pos & (osb->s_clustersize - 1);
1241 if ((cluster_off + local_len) > osb->s_clustersize)
1242 local_len = osb->s_clustersize - cluster_off;
1244 ret = ocfs2_write_cluster(mapping, &desc->c_phys,
1246 desc->c_clear_unwritten,
1249 wc, desc->c_cpos, pos, local_len);
1265 * ocfs2_write_end() wants to know which parts of the target page it
1266 * should complete the write on. It's easiest to compute them ahead of
1267 * time when a more complete view of the write is available.
1269 static void ocfs2_set_target_boundaries(struct ocfs2_super *osb,
1270 struct ocfs2_write_ctxt *wc,
1271 loff_t pos, unsigned len, int alloc)
1273 struct ocfs2_write_cluster_desc *desc;
1275 wc->w_target_from = pos & (PAGE_SIZE - 1);
1276 wc->w_target_to = wc->w_target_from + len;
1282 * Allocating write - we may have different boundaries based
1283 * on page size and cluster size.
1285 * NOTE: We can no longer compute one value from the other as
1286 * the actual write length and user provided length may be
1290 if (wc->w_large_pages) {
1292 * We only care about the 1st and last cluster within
1293 * our range and whether they should be zero'd or not. Either
1294 * value may be extended out to the start/end of a
1295 * newly allocated cluster.
1297 desc = &wc->w_desc[0];
1298 if (desc->c_needs_zero)
1299 ocfs2_figure_cluster_boundaries(osb,
1304 desc = &wc->w_desc[wc->w_clen - 1];
1305 if (desc->c_needs_zero)
1306 ocfs2_figure_cluster_boundaries(osb,
1311 wc->w_target_from = 0;
1312 wc->w_target_to = PAGE_SIZE;
1317 * Check if this extent is marked UNWRITTEN by direct io. If so, we need not to
1318 * do the zero work. And should not to clear UNWRITTEN since it will be cleared
1319 * by the direct io procedure.
1320 * If this is a new extent that allocated by direct io, we should mark it in
1321 * the ip_unwritten_list.
1323 static int ocfs2_unwritten_check(struct inode *inode,
1324 struct ocfs2_write_ctxt *wc,
1325 struct ocfs2_write_cluster_desc *desc)
1327 struct ocfs2_inode_info *oi = OCFS2_I(inode);
1328 struct ocfs2_unwritten_extent *ue = NULL, *new = NULL;
1331 if (!desc->c_needs_zero)
1335 spin_lock(&oi->ip_lock);
1336 /* Needs not to zero no metter buffer or direct. The one who is zero
1337 * the cluster is doing zero. And he will clear unwritten after all
1338 * cluster io finished. */
1339 list_for_each_entry(ue, &oi->ip_unwritten_list, ue_ip_node) {
1340 if (desc->c_cpos == ue->ue_cpos) {
1341 BUG_ON(desc->c_new);
1342 desc->c_needs_zero = 0;
1343 desc->c_clear_unwritten = 0;
1348 if (wc->w_type != OCFS2_WRITE_DIRECT)
1352 spin_unlock(&oi->ip_lock);
1353 new = kmalloc(sizeof(struct ocfs2_unwritten_extent),
1361 /* This direct write will doing zero. */
1362 new->ue_cpos = desc->c_cpos;
1363 new->ue_phys = desc->c_phys;
1364 desc->c_clear_unwritten = 0;
1365 list_add_tail(&new->ue_ip_node, &oi->ip_unwritten_list);
1366 list_add_tail(&new->ue_node, &wc->w_unwritten_list);
1367 wc->w_unwritten_count++;
1370 spin_unlock(&oi->ip_lock);
1377 * Populate each single-cluster write descriptor in the write context
1378 * with information about the i/o to be done.
1380 * Returns the number of clusters that will have to be allocated, as
1381 * well as a worst case estimate of the number of extent records that
1382 * would have to be created during a write to an unwritten region.
1384 static int ocfs2_populate_write_desc(struct inode *inode,
1385 struct ocfs2_write_ctxt *wc,
1386 unsigned int *clusters_to_alloc,
1387 unsigned int *extents_to_split)
1390 struct ocfs2_write_cluster_desc *desc;
1391 unsigned int num_clusters = 0;
1392 unsigned int ext_flags = 0;
1396 *clusters_to_alloc = 0;
1397 *extents_to_split = 0;
1399 for (i = 0; i < wc->w_clen; i++) {
1400 desc = &wc->w_desc[i];
1401 desc->c_cpos = wc->w_cpos + i;
1403 if (num_clusters == 0) {
1405 * Need to look up the next extent record.
1407 ret = ocfs2_get_clusters(inode, desc->c_cpos, &phys,
1408 &num_clusters, &ext_flags);
1414 /* We should already CoW the refcountd extent. */
1415 BUG_ON(ext_flags & OCFS2_EXT_REFCOUNTED);
1418 * Assume worst case - that we're writing in
1419 * the middle of the extent.
1421 * We can assume that the write proceeds from
1422 * left to right, in which case the extent
1423 * insert code is smart enough to coalesce the
1424 * next splits into the previous records created.
1426 if (ext_flags & OCFS2_EXT_UNWRITTEN)
1427 *extents_to_split = *extents_to_split + 2;
1430 * Only increment phys if it doesn't describe
1437 * If w_first_new_cpos is < UINT_MAX, we have a non-sparse
1438 * file that got extended. w_first_new_cpos tells us
1439 * where the newly allocated clusters are so we can
1442 if (desc->c_cpos >= wc->w_first_new_cpos) {
1444 desc->c_needs_zero = 1;
1447 desc->c_phys = phys;
1450 desc->c_needs_zero = 1;
1451 desc->c_clear_unwritten = 1;
1452 *clusters_to_alloc = *clusters_to_alloc + 1;
1455 if (ext_flags & OCFS2_EXT_UNWRITTEN) {
1456 desc->c_clear_unwritten = 1;
1457 desc->c_needs_zero = 1;
1460 ret = ocfs2_unwritten_check(inode, wc, desc);
1474 static int ocfs2_write_begin_inline(struct address_space *mapping,
1475 struct inode *inode,
1476 struct ocfs2_write_ctxt *wc)
1479 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1482 struct ocfs2_dinode *di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1484 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1485 if (IS_ERR(handle)) {
1486 ret = PTR_ERR(handle);
1491 page = find_or_create_page(mapping, 0, GFP_NOFS);
1493 ocfs2_commit_trans(osb, handle);
1499 * If we don't set w_num_pages then this page won't get unlocked
1500 * and freed on cleanup of the write context.
1502 wc->w_pages[0] = wc->w_target_page = page;
1503 wc->w_num_pages = 1;
1505 ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), wc->w_di_bh,
1506 OCFS2_JOURNAL_ACCESS_WRITE);
1508 ocfs2_commit_trans(osb, handle);
1514 if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL))
1515 ocfs2_set_inode_data_inline(inode, di);
1517 if (!PageUptodate(page)) {
1518 ret = ocfs2_read_inline_data(inode, page, wc->w_di_bh);
1520 ocfs2_commit_trans(osb, handle);
1526 wc->w_handle = handle;
1531 int ocfs2_size_fits_inline_data(struct buffer_head *di_bh, u64 new_size)
1533 struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
1535 if (new_size <= le16_to_cpu(di->id2.i_data.id_count))
1540 static int ocfs2_try_to_write_inline_data(struct address_space *mapping,
1541 struct inode *inode, loff_t pos,
1542 unsigned len, struct page *mmap_page,
1543 struct ocfs2_write_ctxt *wc)
1545 int ret, written = 0;
1546 loff_t end = pos + len;
1547 struct ocfs2_inode_info *oi = OCFS2_I(inode);
1548 struct ocfs2_dinode *di = NULL;
1550 trace_ocfs2_try_to_write_inline_data((unsigned long long)oi->ip_blkno,
1551 len, (unsigned long long)pos,
1552 oi->ip_dyn_features);
1555 * Handle inodes which already have inline data 1st.
1557 if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1558 if (mmap_page == NULL &&
1559 ocfs2_size_fits_inline_data(wc->w_di_bh, end))
1560 goto do_inline_write;
1563 * The write won't fit - we have to give this inode an
1564 * inline extent list now.
1566 ret = ocfs2_convert_inline_data_to_extents(inode, wc->w_di_bh);
1573 * Check whether the inode can accept inline data.
1575 if (oi->ip_clusters != 0 || i_size_read(inode) != 0)
1579 * Check whether the write can fit.
1581 di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1583 end > ocfs2_max_inline_data_with_xattr(inode->i_sb, di))
1587 ret = ocfs2_write_begin_inline(mapping, inode, wc);
1594 * This signals to the caller that the data can be written
1599 return written ? written : ret;
1603 * This function only does anything for file systems which can't
1604 * handle sparse files.
1606 * What we want to do here is fill in any hole between the current end
1607 * of allocation and the end of our write. That way the rest of the
1608 * write path can treat it as an non-allocating write, which has no
1609 * special case code for sparse/nonsparse files.
1611 static int ocfs2_expand_nonsparse_inode(struct inode *inode,
1612 struct buffer_head *di_bh,
1613 loff_t pos, unsigned len,
1614 struct ocfs2_write_ctxt *wc)
1617 loff_t newsize = pos + len;
1619 BUG_ON(ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)));
1621 if (newsize <= i_size_read(inode))
1624 ret = ocfs2_extend_no_holes(inode, di_bh, newsize, pos);
1628 /* There is no wc if this is call from direct. */
1630 wc->w_first_new_cpos =
1631 ocfs2_clusters_for_bytes(inode->i_sb, i_size_read(inode));
1636 static int ocfs2_zero_tail(struct inode *inode, struct buffer_head *di_bh,
1641 BUG_ON(!ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)));
1642 if (pos > i_size_read(inode))
1643 ret = ocfs2_zero_extend(inode, di_bh, pos);
1648 int ocfs2_write_begin_nolock(struct address_space *mapping,
1649 loff_t pos, unsigned len, ocfs2_write_type_t type,
1650 struct page **pagep, void **fsdata,
1651 struct buffer_head *di_bh, struct page *mmap_page)
1653 int ret, cluster_of_pages, credits = OCFS2_INODE_UPDATE_CREDITS;
1654 unsigned int clusters_to_alloc, extents_to_split, clusters_need = 0;
1655 struct ocfs2_write_ctxt *wc;
1656 struct inode *inode = mapping->host;
1657 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1658 struct ocfs2_dinode *di;
1659 struct ocfs2_alloc_context *data_ac = NULL;
1660 struct ocfs2_alloc_context *meta_ac = NULL;
1662 struct ocfs2_extent_tree et;
1663 int try_free = 1, ret1;
1666 ret = ocfs2_alloc_write_ctxt(&wc, osb, pos, len, type, di_bh);
1672 if (ocfs2_supports_inline_data(osb)) {
1673 ret = ocfs2_try_to_write_inline_data(mapping, inode, pos, len,
1685 /* Direct io change i_size late, should not zero tail here. */
1686 if (type != OCFS2_WRITE_DIRECT) {
1687 if (ocfs2_sparse_alloc(osb))
1688 ret = ocfs2_zero_tail(inode, di_bh, pos);
1690 ret = ocfs2_expand_nonsparse_inode(inode, di_bh, pos,
1698 ret = ocfs2_check_range_for_refcount(inode, pos, len);
1702 } else if (ret == 1) {
1703 clusters_need = wc->w_clen;
1704 ret = ocfs2_refcount_cow(inode, di_bh,
1705 wc->w_cpos, wc->w_clen, UINT_MAX);
1712 ret = ocfs2_populate_write_desc(inode, wc, &clusters_to_alloc,
1718 clusters_need += clusters_to_alloc;
1720 di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1722 trace_ocfs2_write_begin_nolock(
1723 (unsigned long long)OCFS2_I(inode)->ip_blkno,
1724 (long long)i_size_read(inode),
1725 le32_to_cpu(di->i_clusters),
1726 pos, len, type, mmap_page,
1727 clusters_to_alloc, extents_to_split);
1730 * We set w_target_from, w_target_to here so that
1731 * ocfs2_write_end() knows which range in the target page to
1732 * write out. An allocation requires that we write the entire
1735 if (clusters_to_alloc || extents_to_split) {
1737 * XXX: We are stretching the limits of
1738 * ocfs2_lock_allocators(). It greatly over-estimates
1739 * the work to be done.
1741 ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode),
1743 ret = ocfs2_lock_allocators(inode, &et,
1744 clusters_to_alloc, extents_to_split,
1745 &data_ac, &meta_ac);
1752 data_ac->ac_resv = &OCFS2_I(inode)->ip_la_data_resv;
1754 credits = ocfs2_calc_extend_credits(inode->i_sb,
1756 } else if (type == OCFS2_WRITE_DIRECT)
1757 /* direct write needs not to start trans if no extents alloc. */
1761 * We have to zero sparse allocated clusters, unwritten extent clusters,
1762 * and non-sparse clusters we just extended. For non-sparse writes,
1763 * we know zeros will only be needed in the first and/or last cluster.
1765 if (wc->w_clen && (wc->w_desc[0].c_needs_zero ||
1766 wc->w_desc[wc->w_clen - 1].c_needs_zero))
1767 cluster_of_pages = 1;
1769 cluster_of_pages = 0;
1771 ocfs2_set_target_boundaries(osb, wc, pos, len, cluster_of_pages);
1773 handle = ocfs2_start_trans(osb, credits);
1774 if (IS_ERR(handle)) {
1775 ret = PTR_ERR(handle);
1780 wc->w_handle = handle;
1782 if (clusters_to_alloc) {
1783 ret = dquot_alloc_space_nodirty(inode,
1784 ocfs2_clusters_to_bytes(osb->sb, clusters_to_alloc));
1789 ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), wc->w_di_bh,
1790 OCFS2_JOURNAL_ACCESS_WRITE);
1797 * Fill our page array first. That way we've grabbed enough so
1798 * that we can zero and flush if we error after adding the
1801 ret = ocfs2_grab_pages_for_write(mapping, wc, wc->w_cpos, pos, len,
1802 cluster_of_pages, mmap_page);
1805 * ocfs2_grab_pages_for_write() returns -EAGAIN if it could not lock
1806 * the target page. In this case, we exit with no error and no target
1807 * page. This will trigger the caller, page_mkwrite(), to re-try
1810 if (type == OCFS2_WRITE_MMAP && ret == -EAGAIN) {
1811 BUG_ON(wc->w_target_page);
1820 ret = ocfs2_write_cluster_by_desc(mapping, data_ac, meta_ac, wc, pos,
1828 ocfs2_free_alloc_context(data_ac);
1830 ocfs2_free_alloc_context(meta_ac);
1834 *pagep = wc->w_target_page;
1838 if (clusters_to_alloc)
1839 dquot_free_space(inode,
1840 ocfs2_clusters_to_bytes(osb->sb, clusters_to_alloc));
1842 ocfs2_commit_trans(osb, handle);
1846 * The mmapped page won't be unlocked in ocfs2_free_write_ctxt(),
1847 * even in case of error here like ENOSPC and ENOMEM. So, we need
1848 * to unlock the target page manually to prevent deadlocks when
1849 * retrying again on ENOSPC, or when returning non-VM_FAULT_LOCKED
1852 if (wc->w_target_locked)
1853 unlock_page(mmap_page);
1855 ocfs2_free_write_ctxt(inode, wc);
1858 ocfs2_free_alloc_context(data_ac);
1862 ocfs2_free_alloc_context(meta_ac);
1866 if (ret == -ENOSPC && try_free) {
1868 * Try to free some truncate log so that we can have enough
1869 * clusters to allocate.
1873 ret1 = ocfs2_try_to_free_truncate_log(osb, clusters_need);
1884 static int ocfs2_write_begin(struct file *file, struct address_space *mapping,
1885 loff_t pos, unsigned len,
1886 struct page **pagep, void **fsdata)
1889 struct buffer_head *di_bh = NULL;
1890 struct inode *inode = mapping->host;
1892 ret = ocfs2_inode_lock(inode, &di_bh, 1);
1899 * Take alloc sem here to prevent concurrent lookups. That way
1900 * the mapping, zeroing and tree manipulation within
1901 * ocfs2_write() will be safe against ->read_folio(). This
1902 * should also serve to lock out allocation from a shared
1905 down_write(&OCFS2_I(inode)->ip_alloc_sem);
1907 ret = ocfs2_write_begin_nolock(mapping, pos, len, OCFS2_WRITE_BUFFER,
1908 pagep, fsdata, di_bh, NULL);
1919 up_write(&OCFS2_I(inode)->ip_alloc_sem);
1922 ocfs2_inode_unlock(inode, 1);
1927 static void ocfs2_write_end_inline(struct inode *inode, loff_t pos,
1928 unsigned len, unsigned *copied,
1929 struct ocfs2_dinode *di,
1930 struct ocfs2_write_ctxt *wc)
1934 if (unlikely(*copied < len)) {
1935 if (!PageUptodate(wc->w_target_page)) {
1941 kaddr = kmap_atomic(wc->w_target_page);
1942 memcpy(di->id2.i_data.id_data + pos, kaddr + pos, *copied);
1943 kunmap_atomic(kaddr);
1945 trace_ocfs2_write_end_inline(
1946 (unsigned long long)OCFS2_I(inode)->ip_blkno,
1947 (unsigned long long)pos, *copied,
1948 le16_to_cpu(di->id2.i_data.id_count),
1949 le16_to_cpu(di->i_dyn_features));
1952 int ocfs2_write_end_nolock(struct address_space *mapping,
1953 loff_t pos, unsigned len, unsigned copied, void *fsdata)
1956 unsigned from, to, start = pos & (PAGE_SIZE - 1);
1957 struct inode *inode = mapping->host;
1958 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1959 struct ocfs2_write_ctxt *wc = fsdata;
1960 struct ocfs2_dinode *di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1961 handle_t *handle = wc->w_handle;
1962 struct page *tmppage;
1964 BUG_ON(!list_empty(&wc->w_unwritten_list));
1967 ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode),
1968 wc->w_di_bh, OCFS2_JOURNAL_ACCESS_WRITE);
1976 if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1977 ocfs2_write_end_inline(inode, pos, len, &copied, di, wc);
1978 goto out_write_size;
1981 if (unlikely(copied < len) && wc->w_target_page) {
1982 if (!PageUptodate(wc->w_target_page))
1985 ocfs2_zero_new_buffers(wc->w_target_page, start+copied,
1988 if (wc->w_target_page)
1989 flush_dcache_page(wc->w_target_page);
1991 for(i = 0; i < wc->w_num_pages; i++) {
1992 tmppage = wc->w_pages[i];
1994 /* This is the direct io target page. */
1995 if (tmppage == NULL)
1998 if (tmppage == wc->w_target_page) {
1999 from = wc->w_target_from;
2000 to = wc->w_target_to;
2002 BUG_ON(from > PAGE_SIZE ||
2007 * Pages adjacent to the target (if any) imply
2008 * a hole-filling write in which case we want
2009 * to flush their entire range.
2015 if (page_has_buffers(tmppage)) {
2016 if (handle && ocfs2_should_order_data(inode)) {
2018 ((loff_t)tmppage->index << PAGE_SHIFT) +
2020 loff_t length = to - from;
2021 ocfs2_jbd2_inode_add_write(handle, inode,
2022 start_byte, length);
2024 block_commit_write(tmppage, from, to);
2029 /* Direct io do not update i_size here. */
2030 if (wc->w_type != OCFS2_WRITE_DIRECT) {
2032 if (pos > i_size_read(inode)) {
2033 i_size_write(inode, pos);
2034 mark_inode_dirty(inode);
2036 inode->i_blocks = ocfs2_inode_sector_count(inode);
2037 di->i_size = cpu_to_le64((u64)i_size_read(inode));
2038 inode->i_mtime = inode->i_ctime = current_time(inode);
2039 di->i_mtime = di->i_ctime = cpu_to_le64(inode->i_mtime.tv_sec);
2040 di->i_mtime_nsec = di->i_ctime_nsec = cpu_to_le32(inode->i_mtime.tv_nsec);
2042 ocfs2_update_inode_fsync_trans(handle, inode, 1);
2045 ocfs2_journal_dirty(handle, wc->w_di_bh);
2048 /* unlock pages before dealloc since it needs acquiring j_trans_barrier
2049 * lock, or it will cause a deadlock since journal commit threads holds
2050 * this lock and will ask for the page lock when flushing the data.
2051 * put it here to preserve the unlock order.
2053 ocfs2_unlock_pages(wc);
2056 ocfs2_commit_trans(osb, handle);
2058 ocfs2_run_deallocs(osb, &wc->w_dealloc);
2060 brelse(wc->w_di_bh);
2066 static int ocfs2_write_end(struct file *file, struct address_space *mapping,
2067 loff_t pos, unsigned len, unsigned copied,
2068 struct page *page, void *fsdata)
2071 struct inode *inode = mapping->host;
2073 ret = ocfs2_write_end_nolock(mapping, pos, len, copied, fsdata);
2075 up_write(&OCFS2_I(inode)->ip_alloc_sem);
2076 ocfs2_inode_unlock(inode, 1);
2081 struct ocfs2_dio_write_ctxt {
2082 struct list_head dw_zero_list;
2083 unsigned dw_zero_count;
2085 pid_t dw_writer_pid;
2088 static struct ocfs2_dio_write_ctxt *
2089 ocfs2_dio_alloc_write_ctx(struct buffer_head *bh, int *alloc)
2091 struct ocfs2_dio_write_ctxt *dwc = NULL;
2094 return bh->b_private;
2096 dwc = kmalloc(sizeof(struct ocfs2_dio_write_ctxt), GFP_NOFS);
2099 INIT_LIST_HEAD(&dwc->dw_zero_list);
2100 dwc->dw_zero_count = 0;
2101 dwc->dw_orphaned = 0;
2102 dwc->dw_writer_pid = task_pid_nr(current);
2103 bh->b_private = dwc;
2109 static void ocfs2_dio_free_write_ctx(struct inode *inode,
2110 struct ocfs2_dio_write_ctxt *dwc)
2112 ocfs2_free_unwritten_list(inode, &dwc->dw_zero_list);
2117 * TODO: Make this into a generic get_blocks function.
2119 * From do_direct_io in direct-io.c:
2120 * "So what we do is to permit the ->get_blocks function to populate
2121 * bh.b_size with the size of IO which is permitted at this offset and
2124 * This function is called directly from get_more_blocks in direct-io.c.
2126 * called like this: dio->get_blocks(dio->inode, fs_startblk,
2127 * fs_count, map_bh, dio->rw == WRITE);
2129 static int ocfs2_dio_wr_get_block(struct inode *inode, sector_t iblock,
2130 struct buffer_head *bh_result, int create)
2132 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
2133 struct ocfs2_inode_info *oi = OCFS2_I(inode);
2134 struct ocfs2_write_ctxt *wc;
2135 struct ocfs2_write_cluster_desc *desc = NULL;
2136 struct ocfs2_dio_write_ctxt *dwc = NULL;
2137 struct buffer_head *di_bh = NULL;
2139 unsigned int i_blkbits = inode->i_sb->s_blocksize_bits;
2140 loff_t pos = iblock << i_blkbits;
2141 sector_t endblk = (i_size_read(inode) - 1) >> i_blkbits;
2142 unsigned len, total_len = bh_result->b_size;
2143 int ret = 0, first_get_block = 0;
2145 len = osb->s_clustersize - (pos & (osb->s_clustersize - 1));
2146 len = min(total_len, len);
2149 * bh_result->b_size is count in get_more_blocks according to write
2150 * "pos" and "end", we need map twice to return different buffer state:
2151 * 1. area in file size, not set NEW;
2152 * 2. area out file size, set NEW.
2155 * |--------|---------|---------|---------
2156 * |<-------area in file------->|
2159 if ((iblock <= endblk) &&
2160 ((iblock + ((len - 1) >> i_blkbits)) > endblk))
2161 len = (endblk - iblock + 1) << i_blkbits;
2163 mlog(0, "get block of %lu at %llu:%u req %u\n",
2164 inode->i_ino, pos, len, total_len);
2167 * Because we need to change file size in ocfs2_dio_end_io_write(), or
2168 * we may need to add it to orphan dir. So can not fall to fast path
2169 * while file size will be changed.
2171 if (pos + total_len <= i_size_read(inode)) {
2173 /* This is the fast path for re-write. */
2174 ret = ocfs2_lock_get_block(inode, iblock, bh_result, create);
2175 if (buffer_mapped(bh_result) &&
2176 !buffer_new(bh_result) &&
2180 /* Clear state set by ocfs2_get_block. */
2181 bh_result->b_state = 0;
2184 dwc = ocfs2_dio_alloc_write_ctx(bh_result, &first_get_block);
2185 if (unlikely(dwc == NULL)) {
2191 if (ocfs2_clusters_for_bytes(inode->i_sb, pos + total_len) >
2192 ocfs2_clusters_for_bytes(inode->i_sb, i_size_read(inode)) &&
2193 !dwc->dw_orphaned) {
2195 * when we are going to alloc extents beyond file size, add the
2196 * inode to orphan dir, so we can recall those spaces when
2197 * system crashed during write.
2199 ret = ocfs2_add_inode_to_orphan(osb, inode);
2204 dwc->dw_orphaned = 1;
2207 ret = ocfs2_inode_lock(inode, &di_bh, 1);
2213 down_write(&oi->ip_alloc_sem);
2215 if (first_get_block) {
2216 if (ocfs2_sparse_alloc(osb))
2217 ret = ocfs2_zero_tail(inode, di_bh, pos);
2219 ret = ocfs2_expand_nonsparse_inode(inode, di_bh, pos,
2227 ret = ocfs2_write_begin_nolock(inode->i_mapping, pos, len,
2228 OCFS2_WRITE_DIRECT, NULL,
2229 (void **)&wc, di_bh, NULL);
2235 desc = &wc->w_desc[0];
2237 p_blkno = ocfs2_clusters_to_blocks(inode->i_sb, desc->c_phys);
2238 BUG_ON(p_blkno == 0);
2239 p_blkno += iblock & (u64)(ocfs2_clusters_to_blocks(inode->i_sb, 1) - 1);
2241 map_bh(bh_result, inode->i_sb, p_blkno);
2242 bh_result->b_size = len;
2243 if (desc->c_needs_zero)
2244 set_buffer_new(bh_result);
2246 if (iblock > endblk)
2247 set_buffer_new(bh_result);
2249 /* May sleep in end_io. It should not happen in a irq context. So defer
2250 * it to dio work queue. */
2251 set_buffer_defer_completion(bh_result);
2253 if (!list_empty(&wc->w_unwritten_list)) {
2254 struct ocfs2_unwritten_extent *ue = NULL;
2256 ue = list_first_entry(&wc->w_unwritten_list,
2257 struct ocfs2_unwritten_extent,
2259 BUG_ON(ue->ue_cpos != desc->c_cpos);
2260 /* The physical address may be 0, fill it. */
2261 ue->ue_phys = desc->c_phys;
2263 list_splice_tail_init(&wc->w_unwritten_list, &dwc->dw_zero_list);
2264 dwc->dw_zero_count += wc->w_unwritten_count;
2267 ret = ocfs2_write_end_nolock(inode->i_mapping, pos, len, len, wc);
2271 up_write(&oi->ip_alloc_sem);
2272 ocfs2_inode_unlock(inode, 1);
2280 static int ocfs2_dio_end_io_write(struct inode *inode,
2281 struct ocfs2_dio_write_ctxt *dwc,
2285 struct ocfs2_cached_dealloc_ctxt dealloc;
2286 struct ocfs2_extent_tree et;
2287 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
2288 struct ocfs2_inode_info *oi = OCFS2_I(inode);
2289 struct ocfs2_unwritten_extent *ue = NULL;
2290 struct buffer_head *di_bh = NULL;
2291 struct ocfs2_dinode *di;
2292 struct ocfs2_alloc_context *data_ac = NULL;
2293 struct ocfs2_alloc_context *meta_ac = NULL;
2294 handle_t *handle = NULL;
2295 loff_t end = offset + bytes;
2296 int ret = 0, credits = 0;
2298 ocfs2_init_dealloc_ctxt(&dealloc);
2300 /* We do clear unwritten, delete orphan, change i_size here. If neither
2301 * of these happen, we can skip all this. */
2302 if (list_empty(&dwc->dw_zero_list) &&
2303 end <= i_size_read(inode) &&
2307 ret = ocfs2_inode_lock(inode, &di_bh, 1);
2313 down_write(&oi->ip_alloc_sem);
2315 /* Delete orphan before acquire i_rwsem. */
2316 if (dwc->dw_orphaned) {
2317 BUG_ON(dwc->dw_writer_pid != task_pid_nr(current));
2319 end = end > i_size_read(inode) ? end : 0;
2321 ret = ocfs2_del_inode_from_orphan(osb, inode, di_bh,
2327 di = (struct ocfs2_dinode *)di_bh->b_data;
2329 ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), di_bh);
2331 /* Attach dealloc with extent tree in case that we may reuse extents
2332 * which are already unlinked from current extent tree due to extent
2333 * rotation and merging.
2335 et.et_dealloc = &dealloc;
2337 ret = ocfs2_lock_allocators(inode, &et, 0, dwc->dw_zero_count*2,
2338 &data_ac, &meta_ac);
2344 credits = ocfs2_calc_extend_credits(inode->i_sb, &di->id2.i_list);
2346 handle = ocfs2_start_trans(osb, credits);
2347 if (IS_ERR(handle)) {
2348 ret = PTR_ERR(handle);
2352 ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
2353 OCFS2_JOURNAL_ACCESS_WRITE);
2359 list_for_each_entry(ue, &dwc->dw_zero_list, ue_node) {
2360 ret = ocfs2_mark_extent_written(inode, &et, handle,
2370 if (end > i_size_read(inode)) {
2371 ret = ocfs2_set_inode_size(handle, inode, di_bh, end);
2376 ocfs2_commit_trans(osb, handle);
2378 up_write(&oi->ip_alloc_sem);
2379 ocfs2_inode_unlock(inode, 1);
2383 ocfs2_free_alloc_context(data_ac);
2385 ocfs2_free_alloc_context(meta_ac);
2386 ocfs2_run_deallocs(osb, &dealloc);
2387 ocfs2_dio_free_write_ctx(inode, dwc);
2393 * ocfs2_dio_end_io is called by the dio core when a dio is finished. We're
2394 * particularly interested in the aio/dio case. We use the rw_lock DLM lock
2395 * to protect io on one node from truncation on another.
2397 static int ocfs2_dio_end_io(struct kiocb *iocb,
2402 struct inode *inode = file_inode(iocb->ki_filp);
2406 /* this io's submitter should not have unlocked this before we could */
2407 BUG_ON(!ocfs2_iocb_is_rw_locked(iocb));
2410 mlog_ratelimited(ML_ERROR, "Direct IO failed, bytes = %lld",
2414 ret = ocfs2_dio_end_io_write(inode, private, offset,
2417 ocfs2_dio_free_write_ctx(inode, private);
2420 ocfs2_iocb_clear_rw_locked(iocb);
2422 level = ocfs2_iocb_rw_locked_level(iocb);
2423 ocfs2_rw_unlock(inode, level);
2427 static ssize_t ocfs2_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
2429 struct file *file = iocb->ki_filp;
2430 struct inode *inode = file->f_mapping->host;
2431 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
2432 get_block_t *get_block;
2435 * Fallback to buffered I/O if we see an inode without
2438 if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
2441 /* Fallback to buffered I/O if we do not support append dio. */
2442 if (iocb->ki_pos + iter->count > i_size_read(inode) &&
2443 !ocfs2_supports_append_dio(osb))
2446 if (iov_iter_rw(iter) == READ)
2447 get_block = ocfs2_lock_get_block;
2449 get_block = ocfs2_dio_wr_get_block;
2451 return __blockdev_direct_IO(iocb, inode, inode->i_sb->s_bdev,
2453 ocfs2_dio_end_io, NULL, 0);
2456 const struct address_space_operations ocfs2_aops = {
2457 .dirty_folio = block_dirty_folio,
2458 .read_folio = ocfs2_read_folio,
2459 .readahead = ocfs2_readahead,
2460 .writepage = ocfs2_writepage,
2461 .write_begin = ocfs2_write_begin,
2462 .write_end = ocfs2_write_end,
2464 .direct_IO = ocfs2_direct_IO,
2465 .invalidate_folio = block_invalidate_folio,
2466 .release_folio = ocfs2_release_folio,
2467 .migratepage = buffer_migrate_page,
2468 .is_partially_uptodate = block_is_partially_uptodate,
2469 .error_remove_page = generic_error_remove_page,