Linux Kernel Networking –

advanced topics (5)

Sockets in the kernel

Rami Rosen
[email protected]
Haifux, August 2009
www.haifux.org
All rights reserved.
 Linux Kernel Networking (5)-
advanced topics
● Note:
● This lecture is a sequel to the following 4
lectures I gave in Haifux:
1) Linux Kernel Networking lecture
– http://www.haifux.org/lectures/172/
– slides:http://www.haifux.org/lectures/172/netLec.pdf
2) Advanced Linux Kernel Networking -
Neighboring Subsystem and IPSec lecture
– http://www.haifux.org/lectures/180/
– slides:http://www.haifux.org/lectures/180/netLec2.pdf
 Linux Kernel Networking (5)-
advanced topics
3) Advanced Linux Kernel Networking -
IPv6 in the Linux Kernel lecture
● http://www.haifux.org/lectures/187/
– Slides: http://www.haifux.org/lectures/187/netLec3.pdf
4) Wireless in Linux
http://www.haifux.org/lectures/206/
– Slides: http://www.haifux.org/lectures/206/wirelessLec.pdf
● Table of contents:
– The socket() system call.
– UDP protocol.
– Control Messages.
– Appendixes.
● Note: All code examples in this lecture refer to
the recent 2.6.30 version of the Linux kernel.
TCP socket UDP Socket
Userspace

Layer 4 (TCP,UDP,SCTP,...)

kernel
Layer 3 (Network layer: IPV4/IPV6)

Layer 2 (MAC layer)

● In user space, we have application, session and presentation
layers(tcp/ip refers to all 3 as application layer)
● creating a socket from user space is done by
the socket() system call:
– int socket (int family, int type, int protocol);
– From man 2 socket:
– RETURN VALUE
– On success, a file descriptor for the new socket is returned.
– For open() system call (for files), we also get a file descriptor
as the return value.
– “Everything is a file” Unix paradigm.
● The first parameter, family, is also sometimes referred to as “domain”.
● The family is PF_INET for IPV4 or PF_INET6 for IPV6.
– The family is PF_PACKET for Packet sockets, which operate
at the device driver layer. (Layer 2).
● pcap library for Linux uses PF_PACKET sockets:
– pcap library is in use by sniffers such as tcpdump.
● Also hostapd uses PF_PACKET sockets:
● (hostapd is a wireless access point management project)
● From hostapd:
– drv->monitor_sock = socket(PF_PACKET, SOCK_RAW,
htons(ETH_P_ALL));
● Type:
– SOCK_STREAM and SOCK_DGRAM are the mostly used
types.
● SOCK_STREAM for TCP, SCTP, BLUETOOTH.
● SOCK_DGRAM for UDP.

● SOCK_RAW for RAW sockets.

● There are cases where protocol can be either

SOCK_STREAM or SOCK_DGRAM; for example, Unix

domain socket (AF_UNIX).
– Protocol:usually 0 ( IPPROTO_IP is 0, see:
include/linux/in.h).
– For SCTP, the protocol is IPPROTO_SCTP:
●
sockfd=socket(AF_INET, SOCK_STREAM,IPPROTO_SCTP);
● For bluetooth/RFCOMM:
● socket(AF_BLUETOOTH, SOCK_STREAM,
BTPROTO_RFCOMM);
● SCTP: Stream Control Transmission Protocol.
● For every socket which is created by a userspace
application, there is a corresponding socket struct and
sock struct in the kernel.
● This system call eventually invokes the sock_create()
method in the kernel.
– An instance of struct socket is created (include/linux/net.h)
– struct socket has only 8 members; struct sock has more than 20,
and is one of the biggest structures in the networking stack. You
can easily be confused between them. So the convention is this:
– sock always refers to struct socket.
– sk always refers to struct sock.
struct sock: (include/net/sock.h)

struct sock {
...

struct socket *ssocket;

}
struct socket (include/linux/net.h)

struct socket {
socket_state state;
short type;

unsigned long flags;

struct fasync_struct *fasync_list;

wait_queue_head_t wait;
struct file *file;

struct sock *sk;

const struct proto_ops *ops;
● The state can be
– SS_FREE
– SS_UNCONNECTED
– SS_CONNECTING
– SS_CONNECTED
– SS_DISCONNECTING
● These states are not layer 4 states (like TCP_ESTABLISHED
or TCP_CLOSE).
● The sk_protocol member of struct sock equals to the third
parameter (protocol) of the socket() system call.
● struct proto_ops (interface of struct socket)
inet_stream_ops inet_dgram_ops inet_sockraw_ops
(i.e., TCP sockets) (i.e., UDP sockets) (i.e., RAW sockets)
.family PF_INET PF_INET PF_INET
.owner THIS_MODULE THIS_MODULE THIS_MODULE
.release inet_release inet_release inet_release
.bind inet_bind inet_bind inet_bind
.connect inet_stream_connect inet_dgram_connect inet_dgram_connect
.socketpair sock_no_socketpair sock_no_socketpair sock_no_socketpair
.accept inet_accept sock_no_accept sock_no_accept
.getname inet_getname inet_getname inet_getname
.poll tcp_poll udp_poll datagram_poll
.ioctl inet_ioctl inet_ioctl inet_ioctl
.listen inet_listen sock_no_listen sock_no_listen
.shutdown inet_shutdown inet_shutdown inet_shutdown
.setsockopt sock_common_setsockopt sock_common_setsockopt sock_common_setsockopt
.getsockopt sock_common_getsockopt sock_common_getsockopt sock_common_getsockopt
.sendmsg tcp_sendmsg inet_sendmsg inet_sendmsg
.recvmsg sock_common_recvmsg sock_common_recvmsg sock_common_recvmsg
.mmap sock_no_mmap sock_no_mmap sock_no_mmap
.sendpage tcp_sendpage inet_sendpage inet_sendpage
.splice_read tcp_splice_read - -
● Note: The inet_dgram_ops and inet_sockraw_ops differ only in
the .poll member:
– in inet_dgram_ops it is udp_poll().
– in inet_sockraw_ops, it is datagram_poll().
● Diagram:
struct inet_sock

struct sock (sk)

struct ip_options *opt;

__u8 tos;
__u8 recverr:1;
__u8 hdrincl:1;
.....

inet_sk(sock *sk) => returns the inet_sock which contains sk

● struct sock has three queues: rx , tx and err.

sk_buff sk_buff .... sk_buff

sk_receive_queue

sk_buff sk_buff .... sk_buff sk_write_queue

sk_buff sk_buff
.... sk_buff
sk_error_queue

● Each queue has a lock (spinlock)

● skb_queue_tail() : Adding to the queue
● skb_dequeue() : removing from the queue
– With MSG_PEEK, this is done in two stages:
– skb_peek()
– __skb_unlink(). (to remove the sk_buff from the
queue).
● For the error queue: sock_queue_err_skb() adds
to its tail (include/net/sock.h). Eventually, it also calls
skb_queue_tail().
● Errors can be ICMP errors or EMSGSIZE errors.
● For more about errors,see APPENDIX F: UDP errors.
 UDP and TCP
● No explicit connection setup is done with UDP.
– In TCP there is a preliminary connection setup.
● Packets can be lost in UDP (there is no
retransmission mechanism in the kernel). TCP
on the other hand is reliable (there is a
retransmission mechanism).
● Most of the Internet traffic is TCP (like http,
ssh).
– UDP is for audio/video (RTP)/streaming.
● Note: streaming with VLC is by UDP (RTP).
● Streaming via YouTube is tcp (http).
 The udp header
● There are a very few UDP-based servers like DNS, NTP,
DHCP, TFTP and more.
● For DHCP, it is quite natural to be UDP (Since many times with
DHCP, you don't have a source address, which is a must for TCP).
● TCP implementation is much more complex
– The TCP header is much bigger than UDP header.
The udp header: include/linux/udp.h
struct udphdr {
__be16source;
__be16dest;
__be16len;
__sum16 check;
};
● UDP packet = UDP header + payload
● All members are 2 bytes (16 bits)

source port dest port

len checksum

Payload
 Receiving packets in UDP from
kernel
● UDP kernel sockets can get traffic either from userspace or
from kernel.

USER SPACE
UDP sockets

sock_queue_rcv_skb() UDP – layer 4

KERNEL
ip_local_deliver_finish()
calls udp_rcv() IPv4 - layer 3
NF_INET_LOCAL_IN
hook

Layer 2 (Ethernet)
● From user space, you can receive udp traffic in
three system calls:
– recv() (when the socket is connected)
– recvfrom()
– recvmsg()
● All three are handled by udp_recvmsg() in the kernel.
● Note that fourth parameter of these 3 methods is flags;
however, this parameter is NOT changed upon return. If you are
interested in returned flags , you must use only recvmsg(), and
to retrieve the msg.msg_flags member.
● For example, suppose you have a client-server udp
applications, and the sender sends a packets which is longer
then what the client had allocated for input buffer. The kernel
than truncates the packet, and send MSG_TRUNC flag. In
order to retrieve it, you should use something like:
recvmsg(udpSocket, &msg, flags);
if (msg.msg_flags & MSG_TRUNC)
printf("MSG_TRUNC\n");
● There was a new suggestion recently for
recvmmsg() system call for receiving multiple
messages (By Arnaldo Carvalho de Melo)
● The recvmmsg() will reduce the overhead caused by multiple
system calls of recvmsg() in the usual case.
Receiving packets in UDP from user
space
● UDP kernel sockets can get traffic either from userspace or
from kernel. USER SPACE
UDP sockets
recv() syscall call recvfrom() system call recvmsg() syscall

udp_recvmsg() UDP – layer 4

KERNEL

IPv4 - layer 3
__skb_recv_datagram() :
reads from sk->sk_receive_queue

Layer 2 (Ethernet)
 Receiving packets - udp_rcv()
● udp_rcv() is the handler for all UDP packets
from the IP layer. It handles all incoming
packets in which the protocol field in the ip
header is IPPROTO_UDP (17) after ip layer
finished with them.
See the udp_protocol definition: (net/ipv4/af_inet.c)
struct net_protocol udp_protocol = {
.handler = udp_rcv,
.err_handler = udp_err,
...
};
● In the same way we have :
– raw_rcv() as a handler for raw packets.
– tcp_v4_rcv() as a handler for TCP packets.
– icmp_rcv() as a handler for ICMP packets.
● Kernel implementation: the proto_register()
method registers a protocol handler.
(net/core/sock.c)
udp_rcv() implementation:
● For broadcasts and multicast – there is a special
treatment:
if (rt->rt_flags & (RTCF_BROADCAST|RTCF_MULTICAST))
return __udp4_lib_mcast_deliver(net, skb, uh,
saddr, daddr, udptable);
● Then perform a lookup in a hashtable of struct sock.
– Hash key is created from destination port in the udp header.
– If there is no entry in the hashtable, then there is no sock
listening on this UDP destination port => so send ICMP
back: (of port unreachable).
– icmp_send(skb, ICMP_DEST_UNREACH,
ICMP_PORT_UNREACH, 0);
 udp_rcv()
● In this case, a corresponding SNMP MIB
counter is incremented
(UDP_MIB_NOPORTS).
● UDP_INC_STATS_BH(net, UDP_MIB_NOPORTS, proto ==
IPPROTO_UDPLITE);

● You can see it by:

netstat -s
.....
Udp:
...
35 packets to unknown port received.
 udp_rcv() - contd
● Or, by:
● cat /proc/net/snmp | grep Udp:
Udp: InDatagrams NoPorts InErrors
OutDatagrams RcvbufErrors SndbufErrors
Udp: 14 35 0 30 0 0
● If there is a sock listening on the destination
port, call udp_queue_rcv_skb().
– Eventually calls sock_queue_rcv_skb().
● Which adds the packet to the sk_receive_queue by
skb_queue_tail()
 udp_rcv() diagram
.
udp_rcv()

Multicast
__udp4_lib_rcv __udp4_lib_mcast_deliver

Unicast
icmp_send()
ICMP_DEST_UNREACH,
__udp4_lib_lookup_skb ICMP_PORT_UNREACH

Find a sock in udptable Don't find a sock

udp_queue_rcv_skb sock_queue_rcv_skb
● udp_recvmsg():
● Calls __skb_recv_datagram() , for receiving
one sk_buff.
– The __skb_recv_datagram() may block.
– Eventually, what __skb_recv_datagram() does is
read one sk_buff from the sk_receive_queue
queue.
● memcpy_toiovec() performs the actual copy to
user space by invoking copy_to_user().
● One of the parameters of udp_recvmsg() is a
pointer to struct msghdr. Let's take a look:
 MSGHDR

From include/linux/socket.h:

struct msghdr {
void *msg_name; /* Socket name */
int msg_namelen; /* Length of name */
struct iovec *msg_iov; /* Data blocks */
__kernel_size_t msg_iovlen; /* Number of blocks */
void *msg_control;
__kernel_size_t msg_controllen; /* Length of cmsg list */
unsigned msg_flags;
};
 Control messages (ancillary
messages)
● The msg_control member of msgdhr represent a control
message.
– Sometimes you need to perform some special
things. For example, getting to know what was the
destination address of a received packet.
● Sometimes there is more than one address on a machine
(and also you can have multiple addresses on the same
nic).
– How can we know the destination address of the ip
header in the application?
– struct cmsghdr (/usr/include/bits/socket.h)
represents a control message.
● cmsghdr members can mean different things based on the type
of socket.
● There is a set of macros for handling cmsghdr like
CMSG_FIRSTHDR(), CMSG_NXTHDR(), CMSG_DATA(),
CMSG_LEN() and more.
● There are no control messages for TCP sockets.
 Socket options:
In order to tell the socket to get the information about the packet
destination, we should call setsockopt().
● setsockopt() and getsockopt() - set and get options on a socket.
– Both methods return 0 on success and -1 on error.
● Prototype: int setsockopt(int sockfd, int level, int optname,...
There are two levels of socket options:
To manipulate options at the sockets API level: SOL_SOCKET
To manipulate options at a protocol level, that protocol number
should be used;
– for example, for UDP it is IPPROTO_UDP or SOL_UDP
(both are equal 17) ; see include/linux/in.h and include/linux/socket.h
● SOL_IP is 0.
● There are currently 19 Linux socket options and one
another on option for BSD compatibility.
See Appendix B for a full list of socket options.
● There is an option called IP_PKTINFO.
– We will set the IP_PKTINFO option on a socket in the
following example.
// from /usr/include/bits/in.h
#define IP_PKTINFO 8 /* bool */

/* Structure used for IP_PKTINFO. */

struct in_pktinfo
{
int ipi_ifindex; /* Interface index */
struct in_addr ipi_spec_dst; /* Routing destination address */
struct in_addr ipi_addr; /* Header destination address */
};
const int on = 1;
sockfd = socket(AF_INET, SOCK_DGRAM,0);
if (setsockopt(sockfd, SOL_IP, IP_PKTINFO, &on,
sizeof(on))<0)
perror("setsockopt");
...
...
...
When calling recvmsg(), we will parse the msghr like this:
for (cmptr=CMSG_FIRSTHDR(&msg); cmptr!=NULL;
cmptr=CMSG_NXTHDR(&msg,cmptr))
{
if (cmptr->cmsg_level == SOL_IP && cmptr->cmsg_type ==
IP_PKTINFO)
{
pktinfo = (struct in_pktinfo*)CMSG_DATA(cmptr);
printf("destination=%s\n", inet_ntop(AF_INET, &pktinfo->ipi_addr,
str, sizeof(str)));
}
}
● In the kernel, this calls ip_cmsg_recv() in
net/ipv4/ip_sockglue.c. (which eventually calls
ip_cmsg_recv_pktinfo()).
● You can in this way retrieve other fields of the ip
header:
– For getting the TTL:
● setsockopt(sockfd, SOL_IP, IP_RECVTTL, &on,
sizeof(on))<0).
● But: cmsg_type == IP_TTL.
– For getting ip_options:
● setsockopt() with IP_OPTIONS.
● Note: you cannot get/set ip_options in Java
app.
 Sending packets in UDP
● From user space, you can send udp traffic with
three system calls:
– send() (when the socket is connected).
– sendto()
– sendmsg()
● All three are handled by udp_sendmsg() in the kernel.
● udp_sendmsg() is much simpler than the tcp parallel
method , tcp_sendmsg().
● udp_sendpage() is called when user space calls
sendfile() (to copy a file into a udp socket).
– sendfile() can be used also to copy data between one file
descriptor and another.
– udp_sendpage() invokes udp_sendmsg().
● udp_sendpage() will work only if the nic supports
Scatter/Gather (NETIF_F_SG feature is supported).
 Example – udp client
#include <stdio.h>
#include <arpa/inet.h>
#include <sys/socket.h>
#include <string.h>

int main()
{
int s;
struct sockaddr_in target;
int res;
char buf[10];
target.sin_family = AF_INET;
target.sin_port=htons(999);
inet_aton("192.168.0.121",&target.sin_addr);
strcpy(buf,"message 1:");
s = socket(AF_INET, SOCK_DGRAM, 0);
if (s<0)
perror("socket");
res = sendto(s, buf, sizeof(buf), 0,(struct sockaddr*)&target,
sizeof(struct sockaddr_in));
if (res<0)
perror("sendto");
else
printf("%d bytes were sent\n",res);
}
● For comparison, there is a tcp client in appendix C
● The source port of the UDP packet here is
chosen randomly in the kernel.
● If I want to send from a specified port ?
You can bind to a specific source port (888 in this example) by
adding:
source.sin_family = AF_INET;
source.sin_port = htons(888);
source.sin_addr.s_addr = htonl(INADDR_ANY);
if (bind(s, (struct sockaddr*)&source, sizeof(struct
sockaddr_in)) == -1)
perror("bind");
● You cannot bind to privileged ports (ports lower
than 1024) when you are not root !
– Trying to do this will give:
– “Permission denied” (EPERM).
– You can enable non root binding on privileged port
by running as root: (You will need at least a 2.6.24
kernel)
– setcap 'cap_net_bind_service=+ep' udpclient
– This sets the CAP_NET_BIND_SERVICE
capability.
● You cannot bind on a port which is already
bound.
– Trying to do this will give:
– “Address already in use” (EADDRINUSE)
● You cannot bind twice or more with the same
UDP socket (even if you change the port).
– You will get “bind: Invalid argument” error in such
case (EINVAL)
● If you try connect() on an unbound UDP socket
and then bind() you will also get the EINVAL
error. The reason is that connecting to an
unbound socket will call inet_autobind() to
automatically bind an unbound socket (on a
random port). So after connect(), the socket is
bounded. And the calling bind() again will fail
with EINVAL (since the socket is already
bonded).
● Binding in the kernel for UDP is implemented in
inet_bind() and inet_autobind()
– (in IPV6: inet6_bind() )
 Non local bind
● What happens if we try to bind on a non local address ? (a non
local address can be for example, an address of interface which
is temporarily down)
– We get EADDRNOTAVAIL error:
– “bind: Cannot assign requested address.”
– However, if we set
/proc/sys/net/ipv4/ip_nonlocal_bind to 1, by
– echo "1" > /proc/sys/net/ipv4/ip_nonlocal_bind
– Or adding in /etc/sysctl.conf:
net.ipv4.ip_nonlocal_bind=1
– The bind() will succeed, but it may sometimes break
applications.
● What will happen if in the above udp client example, we will try
setting a broadcast address as the destination (instead of
192.168.0.121), thus:
inet_aton("255.255.255.255",&target.sin_addr);
● We will get EACCESS error (“Permission denied”) for sendto().
● In order that UDP broadcast will work, we have to add:
int flag = 1;
if (setsockopt (s, SOL_SOCKET, SO_BROADCAST,&flag,
sizeof(flag)) < 0)
perror("setsockopt");
 UDP socket options
● For IPPROTO_UDP/SOL_UDP level, we have
two socket options:
● UDP_CORK socket option.
– Added in Linux kernel 2.5.44.
int state=1;
setsockopt(s, IPPROTO_UDP, UDP_CORK, &state,
sizeof(state));
for (j=1;j<1000;j++)
sendto(s,buf1,...)
state=0;
setsockopt(s, IPPROTO_UDP, UDP_CORK, &state,
sizeof(state));
● The above code fragment will call
udp_sendmsg() 1000 times without actually
sending anything on the wire (in the usual case,
when without setsockopt() with UDP_CORK,
1000 packets will be send).
● Only after the second setsockopt() is called,
with UDP_CORK and state=0, one packet is
sent on the wire.
● Kernel implementation: when using
UDP_CORK, udp_sendmsg() passes
MSG_MORE to ip_append_data().
 – Implementation detail: UDP_CORK is not in glibc-header
(/usr/include/netinet/udp.h); you need to add in your
program:
– #define UDP_CORK 1
● UDP_ENCAP socket option.
– For usage with IPSEC.
● Used, for example, in ipsec-tools.
● Note: UDP_ENCAP does not appear yet in the man page
of udp (UDP_CORK does appear).
● Note that there are other socket options at the
SOL_SOCKET level which you can get/set on
UDP sockets: for example, SO_NO_CHECK (to
disable checksum on UDP receive). (see Appendix E).
● SO_DONTROUTE (equivalent to MSG_DONTROUTE in send().
● The SO_DONTROUTE option tells “don't send via a gateway,
only send to directly connected hosts.”
● Adding:
– setsockopt(s, SOL_SOCKET, SO_DONTROUTE, val,
sizeof(one)) < 0)
– And sending the packet to a host on a different network will
cause “Network is unreachable” error to be received.
(ENETUNREACH)
– The same will happen when MSG_DONTROUTE flag is set
in sendto().
● SO_SNDBUF.
● getsockopt(s, SOL_SOCKET, SO_SNDBUF, (void *) &sndbuf).
● Suppose we want to receive ICMP errors with the UDP client
example (like ICMP destination unreachable/port unreachable).
● How can we achieve this ?
● First, we should set this socket option:
– int val=1;
– setsockopt(s, SOL_IP, IP_RECVERR,(char*)&val, sizeof(val));
● Then, we should add a call to a method like this
for receiving error messages:
int recv_err(int s)
{
int res;
char cbuf[512];
struct iovec iov;
struct msghdr msg;
struct cmsghdr *cmsg;
struct sock_extended_err *e;
struct icmphdr icmph;
struct sockaddr_in target;
for (;;)
{
iov.iov_base = &icmph;
iov.iov_len = sizeof(icmph);
msg.msg_name = (void*)&target;
msg.msg_namelen = sizeof(target);
msg.msg_iov = &iov;
msg.msg_iovlen = 1;
msg.msg_flags = 0;
msg.msg_control = cbuf;
msg.msg_controllen = sizeof(cbuf);
res = recvmsg(s, &msg, MSG_ERRQUEUE | MSG_WAITALL);
 if (res<0)
continue;
for (cmsg = CMSG_FIRSTHDR(&msg);cmsg; cmsg =CMSG_NXTHDR(&msg, cmsg))
{
if (cmsg->cmsg_level == SOL_IP)
if (cmsg->cmsg_type == IP_RECVERR)
{
printf("got IP_RECVERR message\n");
e = (struct sock_extended_err*)CMSG_DATA(cmsg);
if (e)
if (e->ee_origin == SO_EE_ORIGIN_ICMP) {
struct sockaddr_in *sin = (struct sockaddr_in *)(e+1);
 if ( (e->ee_type == ICMP_DEST_UNREACH) && (e->ee_code ==
ICMP_PORT_UNREACH) )
printf("Destination port unreachable\n");
}
}
}
}
}
 udp_sendmsg()
● udp_sendmsg(struct kiocb *iocb, struct sock
*sk, struct msghdr *msg, size_t len)
● Sanity checks in udp_sendmsg():
● The destination UDP port must not be 0.
● If we try destination port of 0 we get EINVAL
error as a return value of udp_sendmsg()
– The destination UDP is embedded inside the
msghdr parameter (In fact, msg->msg_name
represents a sockaddr_in; sin_port is sockaddr_in
is the destination port number).
● MSG_OOB is the only illegal flag for UDP.
Returns EOPNOTSUPP error if such a flag is
passed. (only permitted to SOCK_STREAM)
● MSG_OOB is also illegal in AF_UNIX.
● OOB stands for “Out Of Band data”.
● The MSG_OOB flag is permitted in TCP.
– It enables sending one byte of data in urgent mode.
– (telnet , “ctrl/c” for example).
● The destination must be either:
– specified in the msghdr (the name field in msghdr).
– Or the socket is connected.
● sk->sk_state == TCP_ESTABLISHED
– Notice that though this is UDP, we use TCP semantics here.
 Sending packets in UDP (contd)
● In case the socket is not connected, we should
find a route to it; this is done by calling
ip_route_output_flow().
● In case it is connected, we use the route from
the sock (sk_dst_cache member of sk, which is
an instance of dst_entry).
– When the connect() system call was invoked,
ip4_datagram_connect() find the route by
ip_route_connect() and set sk->sk_dst_cache in
sk_dst_set()
● Moving the packet to Layer 3 (IP layer) is done
by ip_append_data().
● In TCP, moving the packet to Layer 3 is done
with ip_queue_xmit().
– What's the difference ?
● UDP does not handle fragmentation;
ip_append_data() does handle fragmentation.
– TCP handles fragmentation in layer 4. So no need
for ip_append_data().
● ip_queue_xmit() is (naturally) a simpler method.
● Basically what the udp_sendmsg() method
does is:
● Finds the route for the packet by
ip_route_output_flow()
● Sends the packet with
ip_local_out(skb)
 Asynchronous I/O
● There is support for Asynchronous I/O in UDP
sockets.
● This means that instead of polling to know if
there is data (by select(), for example), the
kernel sends a SIGIO signal in such a case.
● Using Asynchronous I/O UDP in a user space
application is done in three stages:
– 1) Adding a SIGIO signal handler by calling
sigaction() system call
– 2) Calling fcntl() with F_SETOWN and the pid of our
process to tell the process that it is the owner of the
socket (so that SIGIO signals will be delivered to it).
Several processes can access a socket. If we will not call
fcntl() with F_SETOWN, there can be ambiguity as to which
process will get the SIGIO signal. For example, if we call
fork() the owner of the SIGIO is the parent; but we can call,
in the son, fcntl(s,F_SETOWN, getpid()).
– 3) Setting flags: calling fcntl() with F_SETFL and
O_NONBLOCK | FASYNC.
● In the SIGIO handler, we call recvfrom().
● Example:
struct sockaddr_in source;
struct sigaction handler;
source.sin_family = AF_INET;
source.sin_port = htons(888);
source.sin_addr.s_addr = htonl(INADDR_ANY);
servSocket = socket(AF_INET, SOCK_DGRAM, 0);
bind(servSocket,(struct sockaddr*)&source,sizeof(struct
sockaddr_in));
handler.sa_handler = SIGIOHandler;
sigfillset(&handler.sa_mask);
handler.sa_flags = 0;
sigaction(SIGIO, &handler, 0);
fcntl(servSocket,F_SETOWN, getpid());
fcntl(servSocket,F_SETFL, O_NONBLOCK | FASYNC);

● The fcntl() which sets the O_NONBLOCK | FASYNC flags

invokes sock_fasync() in net/socket.c to add the socket.
– The SIGIOHandler() method will be called when there is
data (since a SIGIO signal was generated) ; it should call
recvmsg().
 Appendix B : Socket options
● Socket options by protocol:
IP protocol (SOL_IP) 19 socket options:
IP_TOS IP_TTL
IP_HDRINCL IP_OPTIONS
IP_ROUTER_ALERT IP_RECVOPTS
IP_RETOPTS IP_PKTINFO
IP_PKTOPTIONS IP_MTU_DISCOVER
IP_RECVERR IP_RECVTTL
IP_RECVTOS IP_MTU
IP_FREEBIND IP_IPSEC_POLICY
IP_XFRM_POLICY IP_PASSSEC
IP_TRANSPARENT
Note: For BSD compatibility there is IP_RECVRETOPTS (which is identical to
IP_RETOPTS).
● AF_UNIX:
– SO_PASSCRED for AF_UNIX sockets.
– Note:For historical reasons these socket options are specified with a
SOL_SOCKET type even though they are PF_UNIX specific.
● UDP:
– UDP_CORK (IPPROTO_UDP level).
● RAW:
– ICMP_FILTER
● TCP:
– TCP_CORK
– TCP_DEFER_ACCEPT
– TCP_INFO
– TCP_KEEPCNT
 – TCP_KEEPIDLE
– TCP_KEEPINTVL
– TCP_LINGER2
– TCP_MAXSEG
– TCP_NODELAY
– TCP_QUICKACK
– TCP_SYNCNT
– TCP_WINDOW_CLAMP
● AF_PACKET
– PACKET_ADD_MEMBERSHIP
– PACKET_DROP_MEMBERSHIP
Socket options for socket level:
SO_DEBUG
SO_REUSEADDR
SO_TYPE
SO_ERROR
SO_DONTROUTE
SO_BROADCAST
SO_SNDBUF
SO_RCVBUF
SO_SNDBUFFORCE
SO_RCVBUFFORCE
SO_KEEPALIVE
SO_OOBINLINE
SO_NO_CHECK
SO_PRIORITY
SO_LINGER
SO_BSDCOMPAT
 Appendix C: tcp client
#include <fcntl.h>
#include <stdlib.h>
#include <errno.h>
#include <stdio.h>
#include <string.h>
#include <sys/sendfile.h>
#include <sys/stat.h>
#include <sys/types.h>
#include <unistd.h>
#include <arpa/inet.h>

int main()
{
 tcp client - contd.
struct sockaddr_in sa;
int sd = socket(PF_INET, SOCK_STREAM, 0);
if (sd<0)
printf("error");
memset(&sa, 0, sizeof(struct sockaddr_in));
sa.sin_family = AF_INET;
sa.sin_port = htons(853);
inet_aton("192.168.0.121",&sa.sin_addr);
if (connect(sd, (struct sockaddr*)&sa, sizeof(sa))<0) {
perror("connect");
exit(0);
}
close(sd);
}
 tcp client - contd.
● If on the other side (192.168.0.121 in this example) there is no
TCP server listening on this port (853) you will get this error for
the socket() system call:
– connect: Connection refused.
● You can send data on this socket by adding, for example:
const char *message = "mymessage";
int length;
length = strlen(message)+1;
res = write(sd, message, length);

● write() is the same as send(), but with no flags.

 Appendix D : ICMP options
● These are ICMP options you can set with
setsockopt on RAW ICMP socket: (see
/usr/include/netinet/ip_icmp.h)
ICMP_ECHOREPLY
ICMP_DEST_UNREACH
ICMP_SOURCE_QUENCH
ICMP_REDIRECT
ICMP_ECHO
ICMP_TIME_EXCEEDED
ICMP_PARAMETERPROB
ICMP_TIMESTAMP
ICMP_TIMESTAMPREPLY
ICMP_INFO_REQUEST
ICMP_INFO_REPLY
ICMP_ADDRESS
ICMP_ADDRESSREPLY
APPENDIX E: flags for send/receive
MSG_OOB
MSG_PEEK
MSG_DONTROUTE
MSG_TRYHARD - Synonym for MSG_DONTROUTE for DECnet
MSG_CTRUNC
MSG_PROBE - Do not send. Only probe path f.e. for MTU
MSG_TRUNC
MSG_DONTWAIT - Nonblocking io
MSG_EOR - End of record
MSG_WAITALL - Wait for a full request
MSG_FIN
MSG_SYN
MSG_CONFIRM - Confirm path validity
MSG_RST
MSG_ERRQUEUE - Fetch message from error queue
MSG_NOSIGNAL - Do not generate SIGPIPE
MSG_MORE 0x8000 - Sender will send more.
 Example: set and get an option
● This simple example demonstrates how to set and get an IP layer option:
#include <stdio.h>
#include <arpa/inet.h>
#include <sys/types.h>
#include <sys/socket.h>
#include <string.h>
int main()
{
int s;
int opt;
int res;
int one = 1;
int size = sizeof(opt);
s = socket(AF_INET, SOCK_DGRAM, 0);
if (s<0)
perror("socket");
res = setsockopt(s, SOL_IP, IP_RECVERR, &one, sizeof(one));
if (res==-1)
perror("setsockopt");
res = getsockopt(s, SOL_IP, IP_RECVERR,&opt,&size);
if (res==-1)
perror("getsockopt");
printf("opt = %d\n",opt);
close(s);
}
 Example: record route option
● This example shows how to send a record route
option.

#define NROUTES 9

int main()
{
int s;
int optlen=0;
struct sockaddr_in target;
int res;
char rspace[3+4*NROUTES+1];
char buf[10];
target.sin_family = AF_INET;
target.sin_port=htons(999);
inet_aton("194.90.1.5",&target.sin_addr);
strcpy(buf,"message 1:");
s = socket(AF_INET, SOCK_DGRAM, 0);
if (s<0)
perror("socket");
memset(rspace, 0, sizeof(rspace));
rspace[0] = IPOPT_NOP;
rspace[1+IPOPT_OPTVAL] = IPOPT_RR;
rspace[1+IPOPT_OLEN] = sizeof(rspace)-1;
rspace[1+IPOPT_OFFSET] = IPOPT_MINOFF;
optlen=40;
if (setsockopt(s, IPPROTO_IP, IP_OPTIONS, rspace,
sizeof(rspace))<0)
{
perror("record route\n");
exit(2);
}
 APPENDIX F: UDP errors
Running :
cat /proc/net/snmp | grep Udp:
will give something like:
Udp: InDatagrams NoPorts InErrors OutDatagrams RcvbufErrors
SndbufErrors
Udp: 2625 1 0 2100 0 0
InErrors - (UDP_MIB_INERRORS)
RcvbufErrors – UDP_MIB_RCVBUFERRORS:
– Incremented in __udp_queue_rcv_skb() (net/ipv4/udp.c).
SndbufErrors – (UDP_MIB_SNDBUFERRORS)
– Incremented in udp_sendmsg()
● Another metric:
– cat /proc/net/udp
– The last column in: drops
● Represents sk->sk_drops.
● Incremented in __udp_queue_rcv_skb()
– net/ipv4/udp.c
● When do RcvbufErrors occur ?
– The total number of bytes queued in sk_receive_queue
queue of a socket is sk->sk_rmem_alloc.
– The total allowed memory of a socket is sk->sk_rcvbuf.
● It can be retrieved with getsockopt() using SO_RCVBUF.
● Each time a packet is received, the sk-
>sk_rmem_alloc is incremented by skb->truesize:
– skb->truesize it the size (in bytes) allocated for the data of
the skb plus the size of sk_buff structure itself.
– This incrementation is done in skb_set_owner_r()
...
atomic_add(skb->truesize, &sk->sk_rmem_alloc);
...
– see: include/net/sock.h
● When the packet is freed by kfree_skb(), we decrement sk-
>sk_rmem_alloc by skb->truesize; this is done in
sock_rfree():
● sock_rfree()
...
atomic_sub(skb->truesize, &sk->sk_rmem_alloc);
...
Immediately in the beginning of sock_queue_rcv_skb(), we
have this check:
if (atomic_read(&sk->sk_rmem_alloc) + skb->truesize >=
(unsigned)sk->sk_rcvbuf) {
err = -ENOMEM;
● When returning -ENOMEM, this notifies the
caller to drop the packet.
● This is done in __udp_queue_rcv_skb() method:
static int __udp_queue_rcv_skb(struct sock *sk, struct sk_buff *skb)
{
...
if ((rc = sock_queue_rcv_skb(sk, skb)) < 0) {
/* Note that an ENOMEM error is charged twice */
if (rc == -ENOMEM) {
UDP_INC_STATS_BH(sock_net(sk), UDP_MIB_RCVBUFERRORS,
is_udplite);
atomic_inc(&sk->sk_drops);
● The default size of sk->sk_rcvbuf is SK_RMEM_MAX
(sysctl_rmem_max).
● It equals to
● (sizeof(struct sk_buff) + 256) * 256
● See: SK_RMEM_MAX definition in
net/core/sock.c
● This can be viewed and modified by:
– /proc/sys/net/core/rmem_default entry.
– getsockopt()/setsockopt() with SO_RCVBUF.
● For the send queue (sk_write_queue), we have in
ip_append_data() a call to sock_alloc_send_skb(), which
eventually invokes sock_alloc_send_pskb().
● In sock_alloc_send_pskb(), we peform this check:
...
if (atomic_read(&sk->sk_wmem_alloc) < sk->sk_sndbuf)
...
● If it is true, everything is fine.
● If not, we end with setting SOCK_ASYNC_NOSPACE and
SOCK_NOSPACE flags of the socket:
set_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
● In udp_sendmsg(), we check the SOCK_NOSPACE flag. If it is
set, we increment the UDP_MIB_SNDBUFERRORS counter.
● sock_alloc_send_pskb() calls skb_set_owner_w().
● In skb_set_owner_w(), we have:
...
atomic_add(skb->truesize, &sk->sk_wmem_alloc);
...
When the packet is freed by kfree_skb(), we decrement
sk_wmem_alloc, in sock_wfree() method:

sock_wfree()
...
atomic_sub(skb->truesize, &sk->sk_wmem_alloc);
...
 Tips
● To find out socket used by a process:
● ls -l /proc/[pid]/fd|grep socket|cut -d: -f3|sed 's/\[//;s/\]//'
● The number returned is the inode number of the socket.
● Information about these sockets can be obtained from
– netstat -ae
● After starting a process which creates a socket, you can see
that the inode cache was incremented by one by:
● more /proc/slabinfo | grep sock
● sock_inode_cache 476 485 768 5 1 : tunables 0 0
0 : slabdata 97 97 0
● The first number, 476, is the number of active objects.
 END

● Thank you!
●

– [email protected]