Received-SPF: pass (google.com: domain of linux-nfs-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
From:   Kuniyuki Iwashima <kuniyu@amazon.co.jp>
To:     <wanghai38@huawei.com>
CC:     <Rao.Shoaib@oracle.com>, <anna.schumaker@netapp.com>,
        <ast@kernel.org>, <bfields@fieldses.org>,
        <christian.brauner@ubuntu.com>, <chuck.lever@oracle.com>,
        <cong.wang@bytedance.com>, <davem@davemloft.net>,
        <dsahern@gmail.com>, <edumazet@google.com>,
        <jakub.kicinski@netronome.com>, <jiang.wang@bytedance.com>,
        <jlayton@kernel.org>, <kolga@netapp.com>, <kuba@kernel.org>,
        <kuniyu@amazon.co.jp>, <linux-kernel@vger.kernel.org>,
        <linux-nfs@vger.kernel.org>, <neilb@suse.com>,
        <netdev@vger.kernel.org>, <nicolas.dichtel@6wind.com>,
        <timo@rothenpieler.org>, <tom@talpey.com>,
        <trond.myklebust@hammerspace.com>, <tyhicks@canonical.com>,
        <viro@zeniv.linux.org.uk>, <wenbin.zeng@gmail.com>,
        <willy@infradead.org>
Subject: Re: [PATCH net 1/2] net: Modify unix_stream_connect to not reference count the netns of kernel sockets
Date:   Tue, 28 Sep 2021 21:50:47 +0900
Message-ID: <20210928125047.40359-1-kuniyu@amazon.co.jp>
In-Reply-To: <20210928031440.2222303-2-wanghai38@huawei.com>
References: <20210928031440.2222303-2-wanghai38@huawei.com>
MIME-Version: 1.0
Content-Transfer-Encoding: 8bit
Content-Type: text/plain
Precedence: bulk

From:   Wang Hai <wanghai38@huawei.com>
Date:   Tue, 28 Sep 2021 11:14:39 +0800
> When use-gss-proxy is set to 1, write_gssp() creates a rpc client in
> gssp_rpc_create(), this increases netns refcount by 2, these refcounts
> are supposed to be released in rpcsec_gss_exit_net(), but it will never
> happen because rpcsec_gss_exit_net() is triggered only when netns
> refcount gets to 0, specifically:
>     refcount=0 -> cleanup_net() -> ops_exit_list -> rpcsec_gss_exit_net
> It is a deadlock situation here, refcount will never get to 0 unless
> rpcsec_gss_exit_net() is called. So, in this case, the netns refcount
> should not be increased.
> 
> In this case, kernel_connect()->unix_stream_connect() will take a netns
> refcount. According to commit 26abe14379f8 ("net: Modify sk_alloc to not
> reference count the netns of kernel sockets."), kernel sockets should not
> take the netns refcount, so unix_stream_connect() should not take
> the netns refcount when the sock is a kernel socket either.
> 
> Signed-off-by: Wang Hai <wanghai38@huawei.com>
> ---
>  net/unix/af_unix.c | 3 ++-
>  1 file changed, 2 insertions(+), 1 deletion(-)
> 
> diff --git a/net/unix/af_unix.c b/net/unix/af_unix.c
> index 92345c9bb60c..af6ba67779c8 100644
> --- a/net/unix/af_unix.c
> +++ b/net/unix/af_unix.c
> @@ -1317,7 +1317,8 @@ static int unix_stream_connect(struct socket *sock, struct sockaddr *uaddr,
>  	err = -ENOMEM;
>  
>  	/* create new sock for complete connection */
> -	newsk = unix_create1(sock_net(sk), NULL, 0, sock->type);
> +	newsk = unix_create1(sock_net(sk), NULL,
> +			     !sk->sk_net_refcnt, sock->type);

This patch conflicts with the commit f4bd73b5a950 for now.
https://git.kernel.org/pub/scm/linux/kernel/git/netdev/net.git/commit/?id=f4bd73b5a950

Could you please rebase and resend the patch set?


>  	if (newsk == NULL)
>  		goto out;
>  
> -- 
> 2.17.1