aboutsummaryrefslogtreecommitdiff
diff options
context:
space:
mode:
authorMark Fasheh <mark.fasheh@oracle.com>2007-11-01 11:37:48 -0700
committerMark Fasheh <mark.fasheh@oracle.com>2007-11-06 15:35:29 -0800
commit4e9563fd55ff4479f2b118d0757d121dd0cfc39c (patch)
tree604066cd76db3d40896263d4a24a7fd5d9824eb2
parent9ea2d32f40434589ea0e136373f7d1545afb411f (diff)
ocfs2: fix write() performance regression
On file systems which don't support sparse files, Ocfs2_map_page_blocks() was reading blocks on appending writes. This caused write performance to suffer dramatically. Fix this by detecting an appending write on a nonsparse fs and skipping the read. Signed-off-by: Mark Fasheh <mark.fasheh@oracle.com>
-rw-r--r--fs/ocfs2/aops.c22
1 files changed, 22 insertions, 0 deletions
diff --git a/fs/ocfs2/aops.c b/fs/ocfs2/aops.c
index c69c1b30015..556e34ccb00 100644
--- a/fs/ocfs2/aops.c
+++ b/fs/ocfs2/aops.c
@@ -729,6 +729,27 @@ static void ocfs2_clear_page_regions(struct page *page,
}
/*
+ * Nonsparse file systems fully allocate before we get to the write
+ * code. This prevents ocfs2_write() from tagging the write as an
+ * allocating one, which means ocfs2_map_page_blocks() might try to
+ * read-in the blocks at the tail of our file. Avoid reading them by
+ * testing i_size against each block offset.
+ */
+static int ocfs2_should_read_blk(struct inode *inode, struct page *page,
+ unsigned int block_start)
+{
+ u64 offset = page_offset(page) + block_start;
+
+ if (ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)))
+ return 1;
+
+ if (i_size_read(inode) > offset)
+ return 1;
+
+ return 0;
+}
+
+/*
* Some of this taken from block_prepare_write(). We already have our
* mapping by now though, and the entire write will be allocating or
* it won't, so not much need to use BH_New.
@@ -781,6 +802,7 @@ int ocfs2_map_page_blocks(struct page *page, u64 *p_blkno,
set_buffer_uptodate(bh);
} else if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
!buffer_new(bh) &&
+ ocfs2_should_read_blk(inode, page, block_start) &&
(block_start < from || block_end > to)) {
ll_rw_block(READ, 1, &bh);
*wait_bh++=bh;