Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
From:   Tao Liu <thomas.liu@ucloud.cn>
To:     pshelar@ovn.org
Cc:     dev@openvswitch.org, netdev@vger.kernel.org,
        linux-kernel@vger.kernel.org, i.maximets@ovn.org,
        jean.tourrilhes@hpe.com, kuba@kernel.org, davem@davemloft.net,
        thomas.liu@ucloud.cn, wenxu@ucloud.cn
Subject: [ovs-dev] [PATCH net v2] openvswitch: meter: fix race when getting now_ms.
Date:   Thu, 13 May 2021 21:08:00 +0800
Message-Id: <20210513130800.31913-1-thomas.liu@ucloud.cn>
MIME-Version: 1.0
Content-Transfer-Encoding: 8bit
Precedence: bulk

We have observed meters working unexpected if traffic is 3+Gbit/s
with multiple connections.

now_ms is not pretected by meter->lock, we may get a negative
long_delta_ms when another cpu updated meter->used, then:
    delta_ms = (u32)long_delta_ms;
which will be a large value.

    band->bucket += delta_ms * band->rate;
then we get a wrong band->bucket.

OpenVswitch userspace datapath has fixed the same issue[1] some
time ago, and we port the implementation to kernel datapath.

[1] https://patchwork.ozlabs.org/project/openvswitch/patch/20191025114436.9746-1-i.maximets@ovn.org/

Fixes: 96fbc13d7e77 ("openvswitch: Add meter infrastructure")
Signed-off-by: Tao Liu <thomas.liu@ucloud.cn>
Suggested-by: Ilya Maximets <i.maximets@ovn.org>
---
Changelog:
v2: just set negative long_delta_ms to zero in case of race for meter lock.
v1: make now_ms protected by meter lock.
---
 net/openvswitch/meter.c | 8 ++++++++
 1 file changed, 8 insertions(+)

diff --git a/net/openvswitch/meter.c b/net/openvswitch/meter.c
index 96b524c..896b8f5 100644
--- a/net/openvswitch/meter.c
+++ b/net/openvswitch/meter.c
@@ -611,6 +611,14 @@ bool ovs_meter_execute(struct datapath *dp, struct sk_buff *skb,
 	spin_lock(&meter->lock);
 
 	long_delta_ms = (now_ms - meter->used); /* ms */
+	if (long_delta_ms < 0) {
+		/* This condition means that we have several threads fighting
+		 * for a meter lock, and the one who received the packets a
+		 * bit later wins. Assuming that all racing threads received
+		 * packets at the same time to avoid overflow.
+		 */
+		long_delta_ms = 0;
+	}
 
 	/* Make sure delta_ms will not be too large, so that bucket will not
 	 * wrap around below.
-- 
1.8.3.1