[CRIU] Re: [PATCH 1/2] sockets: Restore unbound inet sockets v3

[Pavel Emelyanov]

On 06/27/2012 10:14 PM, Cyrill Gorcunov wrote:
> On Wed, Jun 27, 2012 at 09:13:19PM +0400, Pavel Emelyanov wrote:
>>> +	/*
>>> +	 * The restored application will pass backlog explicitly
>>> +	 * on a listen call (if needed) so put some sane value here.
>>
>> Some sane is not good. For UDP sockets this is useless, for TCP you can request
>> for TCPINFO sockoption, it will contain the required value and (!) the socket
>> state, thus you won't have to mess with the SO_ACCEPTCONN above.
>>
> How about this update?

Almost perfect. Two comments:

> +	if (sk->proto == IPPROTO_TCP) {
> +		struct tcp_info info;
> +
> +		aux = sizeof(info);
> +		ret = getsockopt(lfd, SOL_TCP, TCP_INFO, &info, &aux);
> +		if (ret) {
> +			pr_perror("Failt to obtain TCP_INFO");
> +			goto err;
> +		}
> +
> +		if (info.tcpi_state != TCP_CLOSE) {
> +			pr_err("Socket state %d obtained but expected %d\n",
> +			       info.tcpi_state, TCP_CLOSE);
> +			goto err;
> +		}
> +
> +		sk->wqlen = info.tcpi_backoff;
> +	} else
> +		sk->wqlen = 16; /* default value */

This is bad default value.

> +	if (should_free_sk)
> +		xfree(sk);

I'd push this socket into hash thus making the BUG_ON(sk->already_dumped) debug ON for them.

> 	Cyrill