f2fs: avoid IO split due to mixed WB_SYNC_ALL and WB_SYNC_NONE

If two threads try to flush dirty pages in different inodes respectively,
f2fs_write_data_pages() will produce WRITE and WRITE_SYNC one at a time,
resulting in a lot of 4KB seperated IOs.

So, this patch gives higher priority to WB_SYNC_ALL IOs and gathers write
IOs with a big WRITE_SYNC'ed bio.

Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>

fs/f2fs/data.c

@@ -1605,7 +1605,9 @@ continue_unlock:
 				last_idx = page->index;
 			}
 
-			if (--wbc->nr_to_write <= 0 &&
+			/* give a priority to WB_SYNC threads */
+			if ((atomic_read(&F2FS_M_SB(mapping)->wb_sync_req) ||
+					--wbc->nr_to_write <= 0) &&
 					wbc->sync_mode == WB_SYNC_NONE) {
 				done = 1;
 				break;
@@ -1662,9 +1664,18 @@ static int f2fs_write_data_pages(struct address_space *mapping,
 
 	trace_f2fs_writepages(mapping->host, wbc, DATA);
 
+	/* to avoid spliting IOs due to mixed WB_SYNC_ALL and WB_SYNC_NONE */
+	if (wbc->sync_mode == WB_SYNC_ALL)
+		atomic_inc(&sbi->wb_sync_req);
+	else if (atomic_read(&sbi->wb_sync_req))
+		goto skip_write;
+
 	blk_start_plug(&plug);
 	ret = f2fs_write_cache_pages(mapping, wbc);
 	blk_finish_plug(&plug);
+
+	if (wbc->sync_mode == WB_SYNC_ALL)
+		atomic_dec(&sbi->wb_sync_req);
 	/*
 	 * if some pages were truncated, we cannot guarantee its mapping->host
 	 * to detect pending bios.

fs/f2fs/f2fs.h

@@ -887,6 +887,9 @@ struct f2fs_sb_info {
 	/* # of allocated blocks */
 	struct percpu_counter alloc_valid_block_count;
 
+	/* writeback control */
+	atomic_t wb_sync_req;			/* count # of WB_SYNC threads */
+
 	/* valid inode count */
 	struct percpu_counter total_valid_inode_count;
 

fs/f2fs/super.c

@@ -1566,6 +1566,8 @@ static void init_sb_info(struct f2fs_sb_info *sbi)
 	for (i = 0; i < NR_COUNT_TYPE; i++)
 		atomic_set(&sbi->nr_pages[i], 0);
 
+	atomic_set(&sbi->wb_sync_req, 0);
+
 	INIT_LIST_HEAD(&sbi->s_list);
 	mutex_init(&sbi->umount_mutex);
 	mutex_init(&sbi->wio_mutex[NODE]);