==================== libnids-1.17 ==================== 1. Introduction 2. IP defragmentation 3. TCP stream assembly 4. A sample application 5. Basic libnids structures and functions 6. Misc useful hacks 1. Introduction Declarations of data structures and functions defined by libnids are gathered in include file "nids.h". An application which uses libnids must include this file and must be linked with libnids.a. An application's function main usually looks this way: main() { application private processing, not related to libnids optional modification of libnids parameters if (!nids_init() ) something's wrong, terminate; registration of callback functions nids_run(); // not reached in normal situation } Another method is mentioned later. 2. IP defragmentation In order to receive all IP packets seen by libnids (including fragmented ones, packets with invalid checksum et cetera) a programmer should define a callback function of the following type void ip_frag_func(struct ip * a_packet, int len) After calling nids_init, this function should be registered with libnids: nids_register_ip_frag(ip_frag_func); Function ip_frag_func will be called from libnids; parameter a_packet will point to a received datagram, len is the packet length. Analogically, in order to receive only packets, which will be accepted by a target host (that is, packets not fragmented or packets assembled from fragments; a header correctness is verified) one should define a callback function void ip_func(struct ip * a_packet) and register it with nids_register_ip(ip_func); 3. TCP stream assembly In order to receive data exchanged in a TCP stream, one must declare a callback function void tcp_callback(struct tcp_stream * ns, void ** param) Structure tcp_stream provides all info on a TCP connection. For instance, it contains two fields of type struct half_stream (named client and server), each of them describing one side of a connection. We'll explain all its fields later. One of tcp_stream field is named nids_state. Behaviour of tcp_callback depends on value of this field. * ns->nids_state==NIDS_JUST_EST In this case, ns describes a connection which has just been established. Tcp_callback must decide if it wishes to be notified in future of arrival of data in this connection. All the connection parameters are available (IP addresses, ports numbers etc). If the connection is interesting, tcp_callback informs libnids which data it wishes to receive (data to client, to server, urgent data to client, urgent data to server). Then the function returns. * ns->nids_state==NIDS_DATA In this case, new data has arrived. Structures half_stream (members of tcp_stream) contain buffers with data. * The following values of nids_state field : + NIDS_CLOSE + NIDS_RESET + NIDS_TIMEOUT mean that the connection has been closed. Tcp_callback should free allocated resources, if any. * ns->nids_state==NIDS_EXITING In this case, libnids is exiting. This is the applications last opportunity to make use of any data left stored in the half_stream buffers. When reading traffic from a capture file rather than the network, libnids may never see a close, reset, or timeout. If the application has unprocessed data (e.g., from using nids_discard(), this allows the application to process it. 4. A sample application Now let's have a look at a simple application, which displays on stderr data exchanged in all TCP connections seen by libnids. #include <sys/types.h> #include <sys/socket.h> #include <netinet/in.h> #include <netinet/in_systm.h> #include <arpa/inet.h> #include <string.h> #include <stdio.h> #include "nids.h" #define int_ntoa(x) inet_ntoa(*((struct in_addr *)&x)) // struct tuple4 contains addresses and port numbers of the TCP connections // the following auxiliary function produces a string looking like // 10.0.0.1,1024,10.0.0.2,23 char * adres (struct tuple4 addr) { static char buf[256]; strcpy (buf, int_ntoa (addr.saddr)); sprintf (buf + strlen (buf), ",%i,", addr.source); strcat (buf, int_ntoa (addr.daddr)); sprintf (buf + strlen (buf), ",%i", addr.dest); return buf; } void tcp_callback (struct tcp_stream *a_tcp, void ** this_time_not_needed) { char buf[1024]; strcpy (buf, adres (a_tcp->addr)); // we put conn params into buf if (a_tcp->nids_state == NIDS_JUST_EST) { // connection described by a_tcp is established // here we decide, if we wish to follow this stream // sample condition: if (a_tcp->addr.dest!=23) return; // in this simple app we follow each stream, so.. a_tcp->client.collect++; // we want data received by a client a_tcp->server.collect++; // and by a server, too a_tcp->server.collect_urg++; // we want urgent data received by a // server #ifdef WE_WANT_URGENT_DATA_RECEIVED_BY_A_CLIENT a_tcp->client.collect_urg++; // if we don't increase this value, // we won't be notified of urgent data // arrival #endif fprintf (stderr, "%s established\n", buf); return; } if (a_tcp->nids_state == NIDS_CLOSE) { // connection has been closed normally fprintf (stderr, "%s closing\n", buf); return; } if (a_tcp->nids_state == NIDS_RESET) { // connection has been closed by RST fprintf (stderr, "%s reset\n", buf); return; } if (a_tcp->nids_state == NIDS_DATA) { // new data has arrived; gotta determine in what direction // and if it's urgent or not struct half_stream *hlf; if (a_tcp->server.count_new_urg) { // new byte of urgent data has arrived strcat(buf,"(urgent->)"); buf[strlen(buf)+1]=0; buf[strlen(buf)]=a_tcp->server.urgdata; write(1,buf,strlen(buf)); return; } // We don't have to check if urgent data to client has arrived, // because we haven't increased a_tcp->client.collect_urg variable. // So, we have some normal data to take care of. if (a_tcp->client.count_new) { // new data for the client hlf = &a_tcp->client; // from now on, we will deal with hlf var, // which will point to client side of conn strcat (buf, "(<-)"); // symbolic direction of data } else { hlf = &a_tcp->server; // analogical strcat (buf, "(->)"); } fprintf(stderr,"%s",buf); // we print the connection parameters // (saddr, daddr, sport, dport) accompanied // by data flow direction (-> or <-) write(2,hlf->data,hlf->count_new); // we print the newly arrived data } return ; } int main () { // here we can alter libnids params, for instance: // nids_params.n_hosts=256; if (!nids_init ()) { fprintf(stderr,"%s\n",nids_errbuf); exit(1); } nids_register_tcp (tcp_callback); nids_run (); return 0; } 5. Basic libnids structures and functions Now it's time for more systematic description of libnids structures. As mentioned, they're all declared in nids.h struct tuple4 // TCP connection parameters { unsigned short source,dest; // client and server port numbers unsigned long saddr,daddr; // client and server IP addresses }; struct half_stream // structure describing one side of a TCP connection { char state; // socket state (ie TCP_ESTABLISHED ) char collect; // if >0, then data should be stored in // "data" buffer; else // data flowing in this direction will be ignored // have a look at samples/sniff.c for an example // how one can use this field char collect_urg; // analogically, determines if to collect urgent // data char * data; // buffer for normal data unsigned char urgdata; // one-byte buffer for urgent data int count; // how many bytes has been appended to buffer "data" // since the creation of a connection int offset; // offset (in data stream) of first byte stored in // the "data" buffer; additional explanations // follow int count_new; // how many bytes were appended to "data" buffer // last (this) time; if == 0, no new data arrived char count_new_urg; // if != 0, new urgent data arrived ... // other fields are auxiliary for libnids }; struct tcp_stream { struct tuple4 addr; // connections params (saddr, daddr, sport, dport) char nids_state; // logical state of the connection struct half_stream client,server; // structures describing client and // server side of the connection ... // other fields are auxiliary for libnids }; In the above sample program function tcp_callback printed data from hlf->data buffer on stderr, and this data was no longer needed. After tcp_callback return, libnids by default frees space occupied by this data. Field hlf->offset will be increased by number of discarded bytes, and new data will be stored at the beginning of "data" buffer. If the above is not the desired behaviour (for instance, data processor needs at least N bytes of input to operate, and so far libnids received count_new<N bytes) one should call function void nids_discard(struct tcp_stream * a_tcp, int num_bytes) before tcp_callback returns. As a result, after tcp_callback return libnids will discard at most num_bytes first bytes from buffer "data" (updating "offset" field accordingly, and moving rest of the data to the beginning of the buffer). If nids_discard function is never called (like in above sample program), buffer hlf->data contains exactly hlf->count_new bytes. Generally, number of bytes in buffer hlf->data equals hlf->count-hlf->offset. Thanks to nids_discard function, a programmer doesn't have to copy received bytes into a separate buffer - hlf->data will always contain as many bytes, as possible. However, often arises a need to maintain auxiliary data structures per each pair (libnids_callback, tcp stream). For instance, if we wish to detect an attack against wu-ftpd (this attack involves creating deep directory on the server), we need to store somewhere current directory of a ftpd daemon. It will be changed by "CWD" instructions sent by ftp client. That's what the second parameter of tcp_callback is for. It is a pointer to a pointer to data private for each (libnids_callback, tcp stream) pair. Typically, one should use it as follows: void tcp_callback_2 (struct tcp_stream * a_tcp, struct conn_param **ptr) { if (a_tcp->nids_state==NIDS_JUST_EST) { struct conn_param * a_conn; if the connection is uninteresting, return; a_conn=malloc of some data structure init of a_conn *ptr=a_conn // this value will be passed to tcp_callback_2 in future // calls increase some of "collect" fields return; } if (a_tcp->nids_state==NIDS_DATA) { struct conn_param *current_conn_param=*ptr; using current_conn_param and the newly received data from the net we search for attack signatures, possibly modyfying current_conn_param return ; } Functions nids_register_tcp and nids_register_ip* can be called arbitrary number of times. Two different functions (similar to tcp_callback) are allowed to follow the same TCP stream (with a certain non-default exception). Libnids parameters can be changed by modyfication of fields of the global variable nids_params, declared as follows: struct nids_prm { int n_tcp_streams; // size of the hash table used for storing structures // tcp_stream; libnis will follow no more than // 3/4 * n_tcp_streams connections simultaneously // default value: 1040. If set to 0, libnids will // not assemble TCP streams. int n_hosts; // size of the hash table used for storing info on // IP defragmentation; default value: 256 char * filename; // capture filename from which to read packets; // file must be in libpcap format and device must // be set to NULL; default value: NULL char * device; // interface on which libnids will listen for packets; // default value == NULL, in which case device will // be determined by call to pcap_lookupdev; special // value of "all" results in libnids trying to // capture packets on all interfaces (this works only // with Linux kernel > 2.2.0 and libpcap >= 0.6.0); // see also doc/LINUX int sk_buff_size; // size of struct sk_buff, a structure defined by // Linux kernel, used by kernel for packets queuing. If // this parameter has different value from // sizeof(struct sk_buff), libnids can be bypassed // by attacking resource managing of libnis (see TEST // file). If you are paranoid, check sizeof(sk_buff) // on the hosts on your network, and correct this // parameter. Default value: 168 int dev_addon; // how many bytes in structure sk_buff is reserved for // information on net interface; if dev_addon==-1, it // will be corrected during nids_init() according to // type of the interface libnids will listen on. // Default value: -1. void (*syslog)(); // see description below the nids_params definition int syslog_level; // if nids_params.syslog==nids_syslog, then this field // determines loglevel used by reporting events by // system daemon syslogd; default value: LOG_ALERT int scan_num_hosts;// size of hash table used for storing info on port // scanning; the number of simultaneuos port // scan attempts libnids will detect. if set to // 0, port scanning detection will be turned // off. Default value: 256. int scan_num_ports;// how many TCP ports has to be scanned from the same // source. Default value: 10. int scan_delay; // with no more than scan_delay milisecond pause // between two ports, in order to make libnids report // portscan attempt. Default value: 3000 void (*no_mem)(); // called when libnids runs out of memory; it should // terminate the current process int (*ip_filter)(struct ip*); // this function is consulted when an IP // packet arrives; if ip_filter returns non-zero, the // packet is processed, else it is discarded. This way // one can monitor traffic directed at selected hosts // only, not entire subnet. Default function // (nids_ip_filter) always returns 1 char *pcap_filter; // filter string to hand to pcap(3). Default is // NULL. be aware that this applies to the // link-layer, so filters like "tcp dst port 23" // will NOT correctly handle fragmented traffic. int promisc; // if non-zero, the device(s) libnids reads packets // from will be put in promiscuous mode. Default: 1 int one_loop_less; // disabled by default; see the explanation } nids_params; The field syslog of nids_params variable by default contains the address of function nids_syslog, declared as: void nids_syslog (int type, int errnum, struct ip *iph, void *data); Function nids_params.syslog is used to report unusual condition, such as port scan attempts, invalid TCP header flags and other. This field should be assigned the address of a custom event logging function. Function nids_syslog (defined in libnids.c) can be an example on how to decode parameters passed to nids_params.syslog. Nids_syslog logs messages to system daemon syslogd, disregarding such things like message rate per second or free disk space (that is why it should be replaced). If one is interested in UDP datagrams, one should declare void udp_callback(struct tuple4 * addr, char * buf, int len, struct ip * iph); and register it with nids_register_udp(udp_callback) Parameter addr contains address info, buf points to data carried by UDP packet, len is the data length, and iph points to the IP packet which contained the UDP packet. The checksum is verified. 6. Misc useful hacks As a nice toy :) function void nids_killtcp(struct tcp_stream * a_tcp) is implemented. It terminates TCP connection described by a_tcp by sending RST segments. _________________________________________________________________ Using nids_run() has one disadvantage - the application becomes totally packets driven. Sometimes it is necessary to perform some task even when no packets arrive. Instead of nids_run(), one can use function int nids_next() It calls pcap_next() instead of pcap_loop, that is it processes only one packet. If no packet is available, the process will sleep. Nids_next() returns 1 on success, 0 on error (nids_errbuf contains appropriate message then). Typically, when using nids_next(), an aplication will sleep in a select() function, with a snooping socket fd present in read fd_set. This fd can be obtained via a call to int nids_getfd() It returns a file descriptor when succeeded and -1 on error ( nids_errbuf is filled then). _________________________________________________________________ The include file nids.h defines the constants NIDS_MAJOR (1) and NIDS_MINOR (17), which can be used to determine in runtime the version of libnids. Nids.h used to define HAVE_NEW_PCAP as well, but since 1.17 it is nonsupported as obsolete. _________________________________________________________________ When watching a fast network, say 100Mb ethernet, it is a good idea to enlarge kernel buffers dedicated for storing packets. In case of linux, one can call rcvbuf=100*1024; setsockopt(nids_getfd(),SOL_SOCKET, SO_RCVBUF, &rcvbuf, sizeof(rcvbuf)); This setsockopt doubles (approximately) the default kernel buffers size. Unfortunately, there seems to be a limit (about 100KB) for buffers allocated this way. PACKET_RX_RING is supposed to allow to specify arbitrary buffer size, but it has not yet been integrated into libpcap (not in 0.7.1). If you know how to enlarge libpcap buffers on other OS, let me know. _________________________________________________________________ Typically, data carried by a tcp stream can be divided into protocol-dependent records (say, lines of input). A tcp callback can receive an amount of data, which contains more then one record. Therefore, a tcp callback should iterate its protocol parsing routine over the whole amount of data received. This adds complexity to the code. If nids_params.one_loop_less is non-zero, libnids behaviour changes slightly. If a callback consumes some (but not all) of newly arrived data, libnids calls it immediately again. Only non-processed data remain in the buffer, and rcv->count_new is decreased appropriately. Thus, a callback can process only one record at the time - libnids will call it again, until no new data remain or no data can be processed. Unfortunately, this behaviour introduces horrible semantics problems in case of 2+ callbacks reading the same half of a tcp stream. Therefore, if nids_params.one_loop_less is non-zero, you are not allowed to attach two or more callbacks to the same half of tcp stream. Unfortunately, the existing interface is unable to propagate the error to the callback - therefore, you must watch it yourself. You have been warned. _________________________________________________________________ Other applications using libnids can be found in "samples" directory.