在套接字上的Linux中减少TCP最大段大小（MSS）

Question

In a special application in which our server needs to update firmware of low-on-resource sensor/tracking devices we encountered a problem in which sometimes data is lost in the remote devices (clients) receiving packets of the new firmware. The connection is TCP/IP over GPRS network. The devices use SIM900 GSM chip as a network interface.

The problems possibly come because of the device receiving too much data. We tried reducing the traffic by sending packages more rarely but sometimes the error still occured.

We contacted the local retailer of the SIM900 chip who is also responsible for giving technical support and possibly contacting the chinese manufacturer (simcom) of the chip. They said that at first we should try to reduce the TCP MSS (Maximum Segment Size) of our connection.

In our server I did the following:

static int
create_master_socket(unsigned short master_port) {

    static struct sockaddr_in master_address;
    int master_socket = socket(AF_INET,SOCK_STREAM,0);
    if(!master_socket) {
            perror("socket");
            throw runtime_error("Failed to create master socket.");
    }

    int tr=1;
    if(setsockopt(master_socket,SOL_SOCKET,SO_REUSEADDR,&tr,sizeof(int))==-1) {
            perror("setsockopt");
            throw runtime_error("Failed to set SO_REUSEADDR on master socket");
    }

    master_address.sin_family = AF_INET;
    master_address.sin_addr.s_addr = INADDR_ANY;
    master_address.sin_port = htons(master_port);
    uint16_t tcp_maxseg;
    socklen_t tcp_maxseg_len = sizeof(tcp_maxseg);
    if(getsockopt(master_socket, IPPROTO_TCP, TCP_MAXSEG, &tcp_maxseg, &tcp_maxseg_len)) {
            log_error << "Failed to get TCP_MAXSEG for master socket. Reason: " << errno;
            perror("getsockopt");
    } else {
            log_info << "TCP_MAXSEG: " << tcp_maxseg;
    }
    tcp_maxseg = 256;
    if(setsockopt(master_socket, IPPROTO_TCP, TCP_MAXSEG, &tcp_maxseg, tcp_maxseg_len)) {
            log_error << "Failed to set TCP_MAXSEG for master socket. Reason: " << errno;
            perror("setsockopt");
    } else {
            log_info << "TCP_MAXSEG: " << tcp_maxseg;
    }
    if(getsockopt(master_socket, IPPROTO_TCP, TCP_MAXSEG, &tcp_maxseg, &tcp_maxseg_len)) {
            log_error << "Failed to get TCP_MAXSEG for master socket. Reason: " << errno;
            perror("getsockopt");
    } else {
            log_info << "TCP_MAXSEG: " << tcp_maxseg;
    }
    if(bind(master_socket, (struct sockaddr*)&master_address,
                            sizeof(master_address))) {
            perror("bind");
            close(master_socket);
            throw runtime_error("Failed to bind master_socket to port");

    }

    return master_socket;
}

Running the above code results in:

I0807 ... main.cpp:267] TCP_MAXSEG: 536
E0807 ... main.cpp:271] Failed to set TCP_MAXSEG for master socket. Reason: 22 setsockopt: Invalid argument
I0807 ... main.cpp:280] TCP_MAXSEG: 536

As you may see, the problem in the second line of the output: setsockopt returns "Invalid argument".

Why does this happen? I read about some constraints in setting TCP_MAXSEG but I did not encounter any report on such a behaviour as this.

Thanks, Dennis

Answer 1

Unless otherwise noted, optval is a pointer to an int.

but you're using a u_int16. I don't see anything saying that this parameter isn't an int.

edit: Yeah, here is the source code and you can see:

637         if (optlen < sizeof(int))
638                 return -EINVAL;

Answer 2

In addition to xaxxon's answer, just wanted to note my experience with trying to force my Linux to send only maximum TCP segments of a certain size (lower than what they normally are):

• The easiest way I found to do so, was to use iptables:

sudo iptables -A INPUT -p tcp --tcp-flags SYN,RST SYN --destination 1.1.1.1 -j TCPMSS --set-mss 200

This overwrites the remote incoming SYN/ACK packet on an outbound connection, and forces the MSS to a specific value.

Note1: You do not see this in wireshark, since wireshark capture before this happens.

Note 2: Iptables does not allow you to -increase- the MSS, just lower it

• Alternatively, I also tried setting the socket option TCP_MAXSEG, like dennis had done. After taking the fix from xaxxon, this also worked.

Note: You should read the MSS value after the connection has been set up. Otherwise it returns the default value, which put me (and dennis) on the wrong track.

Now finally, I also ran into a number of other things:

• I ran into TCP-offloading issues, where despite my MSS being set correctly, the frames being sent were still shown by wireshark as too big. You can disable this feature by : sudo ethtool -K eth0 tx off sg off tso off . This took me a long time to figure out.

• TCP has lots of fancy things like MTU path discovery, which actually try to dynamically increase the MSS. Fun and cool, but confusing obviously. I did not have issues with it though in my tests

Hope this helps someone trying to do the same thing one day.