writeback: fix possible underflow in write bandwidth calculation

From 1ebf33901ecc75d9496862dceb1ef0377980587c Mon Sep 17 00:00:00 2001
From: Tejun Heo <tj@kernel.org>
Date: Mon, 23 Mar 2015 00:08:19 -0400

2f800fbd777b ("writeback: fix dirtied pages accounting on redirty")
introduced account_page_redirty() which reverts stat updates for a
redirtied page, making BDI_DIRTIED no longer monotonically increasing.

bdi_update_write_bandwidth() uses the delta in BDI_DIRTIED as the
basis for bandwidth calculation.  While unlikely, since the above
patch, the newer value may be lower than the recorded past value and
underflow the bandwidth calculation leading to a wild result.

Fix it by subtracing min of the old and new values when calculating
delta.  AFAIK, there hasn't been any report of it happening but the
resulting erratic behavior would be non-critical and temporary, so
it's possible that the issue is happening without being reported.  The
risk of the fix is very low, so tagged for -stable.

Signed-off-by: Tejun Heo <tj@kernel.org>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Jan Kara <jack@suse.cz>
Cc: Wu Fengguang <fengguang.wu@intel.com>
Cc: Greg Thelen <gthelen@google.com>
Fixes: 2f800fbd777b ("writeback: fix dirtied pages accounting on redirty")
Cc: stable@vger.kernel.org
Signed-off-by: Jens Axboe <axboe@fb.com>

diff --git a/mm/page-writeback.c b/mm/page-writeback.c
index b4fd980a93eb5a9fbd74a56464453a24716b7301..644bcb665773f6e53f50fe6599429595b506f0c5 100644 (file)
--- a/mm/page-writeback.c
+++ b/mm/page-writeback.c
@@ -857,8 +857,11 @@ static void bdi_update_write_bandwidth(struct backing_dev_info *bdi,
         *                   bw * elapsed + write_bandwidth * (period - elapsed)
         * write_bandwidth = ---------------------------------------------------
         *                                          period
+        *
+        * @written may have decreased due to account_page_redirty().
+        * Avoid underflowing @bw calculation.
         */
-       bw = written - bdi->written_stamp;
+       bw = written - min(written, bdi->written_stamp);
        bw *= HZ;
        if (unlikely(elapsed > period)) {
                do_div(bw, elapsed);