[JFFS2] Reduce excessive node count for syslog files.

We currently get fairly poor behaviour with files which get many short
writes, such as system logs. This is because we end up with many tiny
data nodes, and the rbtree gets massive. None of these nodes are
actually obsolete, so they are counted as 'clean' space. Eraseblocks can
be entirely full of these nodes (which are REF_NORMAL instead of
REF_PRISTINE), and still they count entirely towards 'used_size' and the
eraseblocks can sit on the clean_list for a long time without being
picked for GC.

One way to alleviate this in the long term is to account REF_NORMAL
space separately from REF_PRISTINE space, rather than counting them both
towards used_size. Then these eraseblocks can be picked for GC and the
offending nodes will be garbage collected.

The short-term fix, though -- which probably makes sense even if we do
eventually implement the above -- is to merge these nodes as they're
written. When we write the last byte in a page, write the _whole_ page.
This obsoletes the earlier nodes in the page _immediately_ and we don't
even need to wait for the garbage collection to do it.

Original implementation from Ferenc Havasi <havasi@inf.u-szeged.hu>
Signed-off-by: David Woodhouse <dwmw2@infradead.org>

1 files changed, 14 insertions, 6 deletions

diff --git a/fs/jffs2/file.c b/fs/jffs2/file.c
index 9f4171213e5..3349db0a786 100644
--- a/fs/jffs2/file.c
+++ b/fs/jffs2/file.c
@@ -215,12 +215,20 @@ static int jffs2_commit_write (struct file *filp, struct page *pg,
 	D1(printk(KERN_DEBUG "jffs2_commit_write(): ino #%lu, page at 0x%lx, range %d-%d, flags %lx\n",
 		  inode->i_ino, pg->index << PAGE_CACHE_SHIFT, start, end, pg->flags));
 
-	if (!start && end == PAGE_CACHE_SIZE) {
-		/* We need to avoid deadlock with page_cache_read() in
-		   jffs2_garbage_collect_pass(). So we have to mark the
-		   page up to date, to prevent page_cache_read() from
-		   trying to re-lock it. */
-		SetPageUptodate(pg);
+	if (end == PAGE_CACHE_SIZE) {
+		if (!start) {
+			/* We need to avoid deadlock with page_cache_read() in
+			   jffs2_garbage_collect_pass(). So we have to mark the
+			   page up to date, to prevent page_cache_read() from
+			   trying to re-lock it. */
+			SetPageUptodate(pg);
+		} else {
+			/* When writing out the end of a page, write out the 
+			   _whole_ page. This helps to reduce the number of
+			   nodes in files which have many short writes, like
+			   syslog files. */
+			start = aligned_start = 0;
+		}
 	}
 
 	ri = jffs2_alloc_raw_inode();