RE: socket leaks observed in Linux kernel's passive close path

From: "Arankal, Nagaraj" <nagaraj.p.arankal@hpe.com>
To: Andrew Lunn <andrew@lunn.ch>
Cc: "netdev@vger.kernel.org" <netdev@vger.kernel.org>
Subject: RE: socket leaks observed in Linux kernel's passive close path
Date: Thu, 13 Oct 2022 14:44:02 +0000	[thread overview]
Message-ID: <SJ0PR84MB184707E40732357494D0EC17B2259@SJ0PR84MB1847.NAMPRD84.PROD.OUTLOOK.COM> (raw)
In-Reply-To: <Y0genWOLGfy2kQ/M@lunn.ch>

Hi Andrew,
Thanks for looking into this,  I have not tested this on V6.0 kernel, and as far as I know I have not observed any fixes in this area, that's why I posted this, as this seems to be a valid case.

Thanks,
Nagaraj P Arankal

-----Original Message-----
From: Andrew Lunn <andrew@lunn.ch> 
Sent: Thursday, October 13, 2022 7:50 PM
To: Arankal, Nagaraj <nagaraj.p.arankal@hpe.com>
Cc: netdev@vger.kernel.org
Subject: Re: socket leaks observed in Linux kernel's passive close path

On Thu, Oct 13, 2022 at 06:47:56AM +0000, Arankal, Nagaraj wrote:
> Description:
> We have observed a strange race condition , where sockets are not freed in kernel in the following condition.
> We have a kernel module , which monitors the TCP connection state changes , as part of the functionality it replaces the default sk_destruct function of all TCP sockets with our module specific routine.  Looks like sk_destruct() is not invoked in following condition and hence the sockets are leaked despite receiving RESET from the remote.
> 
> 1.	Establish a TCP connection between Host A and Host B.
> 2.	Make the client at B to initiate the CLOSE() immediately after 3-way handshake.
> 3.	Server end sends huge amount of data to client and does close on FD.
> 4.	FIN from the client is not ACKED, and server is busy sending the data.
> 5.	RESET is received from the remote client.
> 6.	Sk_destruct() is not invoked due to non-null sk_refcnt or sk_wmem_alloc count.
> 
> Kernel version: Debian Linux 4.19.y(238,247)

Is this reproducible with a modern kernel? v6.0? If this is already fixed, we need to identify what change fixed it, and get it back ported. If it is broken in v6.0, and net-next, it first needs fixing in net-next, and then back porting to the different LTS kernels.

   Andrew