f2fs: avoid high cpu usage in discard thread
authorChao Yu <yuchao0@huawei.com>
Mon, 8 Jan 2018 10:48:33 +0000 (18:48 +0800)
committerJaegeuk Kim <jaegeuk@kernel.org>
Tue, 16 Jan 2018 23:40:00 +0000 (15:40 -0800)
We take very long time to finish generic/476, this is because we will
check consistence of all discard entries in global rb tree while
traversing all different granularity pending lists, even when the list
is empty, in order to avoid that unneeded overhead, we have to skip
the check when coming up an empty list.

generic/476 time consumption:
cost
Before patch & w/o consistence check 57s
Before patch & w/ consistence check 1426s
After patch 78s

Signed-off-by: Chao Yu <yuchao0@huawei.com>
Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
fs/f2fs/segment.c

index b4b9098..35009db 100644 (file)
@@ -1204,6 +1204,8 @@ static int __issue_discard_cmd(struct f2fs_sb_info *sbi,
                pend_list = &dcc->pend_list[i];
 
                mutex_lock(&dcc->cmd_lock);
+               if (list_empty(pend_list))
+                       goto next;
                f2fs_bug_on(sbi, !__check_rb_tree_consistence(sbi, &dcc->root));
                blk_start_plug(&plug);
                list_for_each_entry_safe(dc, tmp, pend_list, list) {
@@ -1222,6 +1224,7 @@ skip:
                                break;
                }
                blk_finish_plug(&plug);
+next:
                mutex_unlock(&dcc->cmd_lock);
 
                if (iter >= dpolicy->max_requests)