ext4: fix performance regression in ext4_writepages

Commit 4e7ea81db5(ext4: restructure writeback path) introduces another
performance regression on random write:

- one more page may be added to ext4 extent in
  mpage_prepare_extent_to_map, and will be submitted for I/O so
  nr_to_write will become -1 before 'done' is set

- the worse thing is that dirty pages may still be retrieved from page
  cache after nr_to_write becomes negative, so lots of small chunks
  can be submitted to block device when page writeback is catching up
  with write path, and performance is hurted.

On one arm A15 board with sata 3.0 SSD(CPU: 1.5GHz dura core, RAM:
2GB, SATA controller: 3.0Gbps), this patch can improve below test's
result from 157MB/sec to 174MB/sec(>10%):

	dd if=/dev/zero of=./z.img bs=8K count=512K

The above test is actually prototype of block write in bonnie++
utility.

This patch makes sure no more pages than nr_to_write can be added to
extent for mapping, so that nr_to_write won't become negative.

Cc: linux-ext4@vger.kernel.org
Acked-by: Jan Kara <jack@suse.cz>
Signed-off-by: Ming Lei <ming.lei@canonical.com>
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>

diff --git a/fs/ext4/inode.c b/fs/ext4/inode.c
index e7e5b3d8f002565150b7352a7e4ec8251211c854..94aac67b55c96b8ac5cf3ba982875361f4917beb 100644 (file)
--- a/fs/ext4/inode.c
+++ b/fs/ext4/inode.c
@@ -2298,6 +2298,7 @@ static int mpage_prepare_extent_to_map(struct mpage_da_data *mpd)
        struct address_space *mapping = mpd->inode->i_mapping;
        struct pagevec pvec;
        unsigned int nr_pages;
+       long left = mpd->wbc->nr_to_write;
        pgoff_t index = mpd->first_page;
        pgoff_t end = mpd->last_page;
        int tag;
@@ -2333,6 +2334,17 @@ static int mpage_prepare_extent_to_map(struct mpage_da_data *mpd)
                        if (page->index > end)
                                goto out;
 
+                       /*
+                        * Accumulated enough dirty pages? This doesn't apply
+                        * to WB_SYNC_ALL mode. For integrity sync we have to
+                        * keep going because someone may be concurrently
+                        * dirtying pages, and we might have synced a lot of
+                        * newly appeared dirty pages, but have not synced all
+                        * of the old dirty pages.
+                        */
+                       if (mpd->wbc->sync_mode == WB_SYNC_NONE && left <= 0)
+                               goto out;
+
                        /* If we can't merge this page, we are done. */
                        if (mpd->map.m_len > 0 && mpd->next_page != page->index)
                                goto out;
@@ -2367,19 +2379,7 @@ static int mpage_prepare_extent_to_map(struct mpage_da_data *mpd)
                        if (err <= 0)
                                goto out;
                        err = 0;
-
-                       /*
-                        * Accumulated enough dirty pages? This doesn't apply
-                        * to WB_SYNC_ALL mode. For integrity sync we have to
-                        * keep going because someone may be concurrently
-                        * dirtying pages, and we might have synced a lot of
-                        * newly appeared dirty pages, but have not synced all
-                        * of the old dirty pages.
-                        */
-                       if (mpd->wbc->sync_mode == WB_SYNC_NONE &&
-                           mpd->next_page - mpd->first_page >=
-                                                       mpd->wbc->nr_to_write)
-                               goto out;
+                       left--;
                }
                pagevec_release(&pvec);
                cond_resched();