f2fs: fix to correct freed section number during gc
authorChao Yu <chao2.yu@samsung.com>
Mon, 28 Sep 2015 09:42:24 +0000 (17:42 +0800)
committerJaegeuk Kim <jaegeuk@kernel.org>
Fri, 9 Oct 2015 23:20:54 +0000 (16:20 -0700)
This patch fixes to maintain the right section count freed in garbage
collecting when triggering a foreground gc.

Besides, when a foreground gc is running on current selected section, once
we fail to gc one segment, it's better to abandon gcing the left segments
in current section, because anyway we will select next victim for
foreground gc, so gc on the left segments in previous section will become
overhead and also cause the long latency for caller.

Signed-off-by: Chao Yu <chao2.yu@samsung.com>
Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
fs/f2fs/gc.c

index 782b8e72c094a8ef3a04b20f7be76e3519931715..b6e03ebc703ca81988e6ed8a941d64f012e2f309 100644 (file)
@@ -802,7 +802,7 @@ int f2fs_gc(struct f2fs_sb_info *sbi)
        unsigned int segno = NULL_SEGNO;
        unsigned int i;
        int gc_type = BG_GC;
-       int nfree = 0;
+       int sec_freed = 0;
        int ret = -1;
        struct cp_control cpc;
        struct gc_inode_list gc_list = {
@@ -817,7 +817,7 @@ gc_more:
        if (unlikely(f2fs_cp_error(sbi)))
                goto stop;
 
-       if (gc_type == BG_GC && has_not_enough_free_secs(sbi, nfree)) {
+       if (gc_type == BG_GC && has_not_enough_free_secs(sbi, sec_freed)) {
                gc_type = FG_GC;
                if (__get_victim(sbi, &segno, gc_type) || prefree_segments(sbi))
                        write_checkpoint(sbi, &cpc);
@@ -832,13 +832,23 @@ gc_more:
                ra_meta_pages(sbi, GET_SUM_BLOCK(sbi, segno), sbi->segs_per_sec,
                                                                META_SSA);
 
-       for (i = 0; i < sbi->segs_per_sec; i++)
-               nfree += do_garbage_collect(sbi, segno + i, &gc_list, gc_type);
+       for (i = 0; i < sbi->segs_per_sec; i++) {
+               /*
+                * for FG_GC case, halt gcing left segments once failed one
+                * of segments in selected section to avoid long latency.
+                */
+               if (!do_garbage_collect(sbi, segno + i, &gc_list, gc_type) &&
+                               gc_type == FG_GC)
+                       break;
+       }
+
+       if (i == sbi->segs_per_sec && gc_type == FG_GC)
+               sec_freed++;
 
        if (gc_type == FG_GC)
                sbi->cur_victim_sec = NULL_SEGNO;
 
-       if (has_not_enough_free_secs(sbi, nfree))
+       if (has_not_enough_free_secs(sbi, sec_freed))
                goto gc_more;
 
        if (gc_type == FG_GC)