xref: /freebsd/contrib/libpcap/pcap-linux.c (revision 1e413cf93298b5b97441a21d9a50fdcd0ee9945e)
1 /*
2  *  pcap-linux.c: Packet capture interface to the Linux kernel
3  *
4  *  Copyright (c) 2000 Torsten Landschoff <torsten@debian.org>
5  *  		       Sebastian Krahmer  <krahmer@cs.uni-potsdam.de>
6  *
7  *  License: BSD
8  *
9  *  Redistribution and use in source and binary forms, with or without
10  *  modification, are permitted provided that the following conditions
11  *  are met:
12  *
13  *  1. Redistributions of source code must retain the above copyright
14  *     notice, this list of conditions and the following disclaimer.
15  *  2. Redistributions in binary form must reproduce the above copyright
16  *     notice, this list of conditions and the following disclaimer in
17  *     the documentation and/or other materials provided with the
18  *     distribution.
19  *  3. The names of the authors may not be used to endorse or promote
20  *     products derived from this software without specific prior
21  *     written permission.
22  *
23  *  THIS SOFTWARE IS PROVIDED ``AS IS'' AND WITHOUT ANY EXPRESS OR
24  *  IMPLIED WARRANTIES, INCLUDING, WITHOUT LIMITATION, THE IMPLIED
25  *  WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE.
26  */
27 
28 #ifndef lint
29 static const char rcsid[] _U_ =
30     "@(#) $Header: /tcpdump/master/libpcap/pcap-linux.c,v 1.110.2.14 2006/10/12 17:26:58 guy Exp $ (LBL)";
31 #endif
32 
33 /*
34  * Known problems with 2.0[.x] kernels:
35  *
36  *   - The loopback device gives every packet twice; on 2.2[.x] kernels,
37  *     if we use PF_PACKET, we can filter out the transmitted version
38  *     of the packet by using data in the "sockaddr_ll" returned by
39  *     "recvfrom()", but, on 2.0[.x] kernels, we have to use
40  *     PF_INET/SOCK_PACKET, which means "recvfrom()" supplies a
41  *     "sockaddr_pkt" which doesn't give us enough information to let
42  *     us do that.
43  *
44  *   - We have to set the interface's IFF_PROMISC flag ourselves, if
45  *     we're to run in promiscuous mode, which means we have to turn
46  *     it off ourselves when we're done; the kernel doesn't keep track
47  *     of how many sockets are listening promiscuously, which means
48  *     it won't get turned off automatically when no sockets are
49  *     listening promiscuously.  We catch "pcap_close()" and, for
50  *     interfaces we put into promiscuous mode, take them out of
51  *     promiscuous mode - which isn't necessarily the right thing to
52  *     do, if another socket also requested promiscuous mode between
53  *     the time when we opened the socket and the time when we close
54  *     the socket.
55  *
56  *   - MSG_TRUNC isn't supported, so you can't specify that "recvfrom()"
57  *     return the amount of data that you could have read, rather than
58  *     the amount that was returned, so we can't just allocate a buffer
59  *     whose size is the snapshot length and pass the snapshot length
60  *     as the byte count, and also pass MSG_TRUNC, so that the return
61  *     value tells us how long the packet was on the wire.
62  *
63  *     This means that, if we want to get the actual size of the packet,
64  *     so we can return it in the "len" field of the packet header,
65  *     we have to read the entire packet, not just the part that fits
66  *     within the snapshot length, and thus waste CPU time copying data
67  *     from the kernel that our caller won't see.
68  *
69  *     We have to get the actual size, and supply it in "len", because
70  *     otherwise, the IP dissector in tcpdump, for example, will complain
71  *     about "truncated-ip", as the packet will appear to have been
72  *     shorter, on the wire, than the IP header said it should have been.
73  */
74 
75 
76 #ifdef HAVE_CONFIG_H
77 #include "config.h"
78 #endif
79 
80 #include "pcap-int.h"
81 #include "sll.h"
82 
83 #ifdef HAVE_DAG_API
84 #include "pcap-dag.h"
85 #endif /* HAVE_DAG_API */
86 
87 #ifdef HAVE_SEPTEL_API
88 #include "pcap-septel.h"
89 #endif /* HAVE_SEPTEL_API */
90 
91 #include <errno.h>
92 #include <stdlib.h>
93 #include <unistd.h>
94 #include <fcntl.h>
95 #include <string.h>
96 #include <sys/socket.h>
97 #include <sys/ioctl.h>
98 #include <sys/utsname.h>
99 #include <net/if.h>
100 #include <netinet/in.h>
101 #include <linux/if_ether.h>
102 #include <net/if_arp.h>
103 
104 /*
105  * If PF_PACKET is defined, we can use {SOCK_RAW,SOCK_DGRAM}/PF_PACKET
106  * sockets rather than SOCK_PACKET sockets.
107  *
108  * To use them, we include <linux/if_packet.h> rather than
109  * <netpacket/packet.h>; we do so because
110  *
111  *	some Linux distributions (e.g., Slackware 4.0) have 2.2 or
112  *	later kernels and libc5, and don't provide a <netpacket/packet.h>
113  *	file;
114  *
115  *	not all versions of glibc2 have a <netpacket/packet.h> file
116  *	that defines stuff needed for some of the 2.4-or-later-kernel
117  *	features, so if the system has a 2.4 or later kernel, we
118  *	still can't use those features.
119  *
120  * We're already including a number of other <linux/XXX.h> headers, and
121  * this code is Linux-specific (no other OS has PF_PACKET sockets as
122  * a raw packet capture mechanism), so it's not as if you gain any
123  * useful portability by using <netpacket/packet.h>
124  *
125  * XXX - should we just include <linux/if_packet.h> even if PF_PACKET
126  * isn't defined?  It only defines one data structure in 2.0.x, so
127  * it shouldn't cause any problems.
128  */
129 #ifdef PF_PACKET
130 # include <linux/if_packet.h>
131 
132  /*
133   * On at least some Linux distributions (for example, Red Hat 5.2),
134   * there's no <netpacket/packet.h> file, but PF_PACKET is defined if
135   * you include <sys/socket.h>, but <linux/if_packet.h> doesn't define
136   * any of the PF_PACKET stuff such as "struct sockaddr_ll" or any of
137   * the PACKET_xxx stuff.
138   *
139   * So we check whether PACKET_HOST is defined, and assume that we have
140   * PF_PACKET sockets only if it is defined.
141   */
142 # ifdef PACKET_HOST
143 #  define HAVE_PF_PACKET_SOCKETS
144 # endif /* PACKET_HOST */
145 #endif /* PF_PACKET */
146 
147 #ifdef SO_ATTACH_FILTER
148 #include <linux/types.h>
149 #include <linux/filter.h>
150 #endif
151 
152 #ifndef __GLIBC__
153 typedef int		socklen_t;
154 #endif
155 
156 #ifndef MSG_TRUNC
157 /*
158  * This is being compiled on a system that lacks MSG_TRUNC; define it
159  * with the value it has in the 2.2 and later kernels, so that, on
160  * those kernels, when we pass it in the flags argument to "recvfrom()"
161  * we're passing the right value and thus get the MSG_TRUNC behavior
162  * we want.  (We don't get that behavior on 2.0[.x] kernels, because
163  * they didn't support MSG_TRUNC.)
164  */
165 #define MSG_TRUNC	0x20
166 #endif
167 
168 #ifndef SOL_PACKET
169 /*
170  * This is being compiled on a system that lacks SOL_PACKET; define it
171  * with the value it has in the 2.2 and later kernels, so that we can
172  * set promiscuous mode in the good modern way rather than the old
173  * 2.0-kernel crappy way.
174  */
175 #define SOL_PACKET	263
176 #endif
177 
178 #define MAX_LINKHEADER_SIZE	256
179 
180 /*
181  * When capturing on all interfaces we use this as the buffer size.
182  * Should be bigger then all MTUs that occur in real life.
183  * 64kB should be enough for now.
184  */
185 #define BIGGER_THAN_ALL_MTUS	(64*1024)
186 
187 /*
188  * Prototypes for internal functions
189  */
190 static void map_arphrd_to_dlt(pcap_t *, int, int);
191 static int live_open_old(pcap_t *, const char *, int, int, char *);
192 static int live_open_new(pcap_t *, const char *, int, int, char *);
193 static int pcap_read_linux(pcap_t *, int, pcap_handler, u_char *);
194 static int pcap_read_packet(pcap_t *, pcap_handler, u_char *);
195 static int pcap_inject_linux(pcap_t *, const void *, size_t);
196 static int pcap_stats_linux(pcap_t *, struct pcap_stat *);
197 static int pcap_setfilter_linux(pcap_t *, struct bpf_program *);
198 static int pcap_setdirection_linux(pcap_t *, pcap_direction_t);
199 static void pcap_close_linux(pcap_t *);
200 
201 /*
202  * Wrap some ioctl calls
203  */
204 #ifdef HAVE_PF_PACKET_SOCKETS
205 static int	iface_get_id(int fd, const char *device, char *ebuf);
206 #endif
207 static int	iface_get_mtu(int fd, const char *device, char *ebuf);
208 static int 	iface_get_arptype(int fd, const char *device, char *ebuf);
209 #ifdef HAVE_PF_PACKET_SOCKETS
210 static int 	iface_bind(int fd, int ifindex, char *ebuf);
211 #endif
212 static int 	iface_bind_old(int fd, const char *device, char *ebuf);
213 
214 #ifdef SO_ATTACH_FILTER
215 static int	fix_program(pcap_t *handle, struct sock_fprog *fcode);
216 static int	fix_offset(struct bpf_insn *p);
217 static int	set_kernel_filter(pcap_t *handle, struct sock_fprog *fcode);
218 static int	reset_kernel_filter(pcap_t *handle);
219 
220 static struct sock_filter	total_insn
221 	= BPF_STMT(BPF_RET | BPF_K, 0);
222 static struct sock_fprog	total_fcode
223 	= { 1, &total_insn };
224 #endif
225 
226 /*
227  *  Get a handle for a live capture from the given device. You can
228  *  pass NULL as device to get all packages (without link level
229  *  information of course). If you pass 1 as promisc the interface
230  *  will be set to promiscous mode (XXX: I think this usage should
231  *  be deprecated and functions be added to select that later allow
232  *  modification of that values -- Torsten).
233  *
234  *  See also pcap(3).
235  */
236 pcap_t *
237 pcap_open_live(const char *device, int snaplen, int promisc, int to_ms,
238     char *ebuf)
239 {
240 	pcap_t		*handle;
241 	int		mtu;
242 	int		err;
243 	int		live_open_ok = 0;
244 	struct utsname	utsname;
245 
246 #ifdef HAVE_DAG_API
247 	if (strstr(device, "dag")) {
248 		return dag_open_live(device, snaplen, promisc, to_ms, ebuf);
249 	}
250 #endif /* HAVE_DAG_API */
251 
252 #ifdef HAVE_SEPTEL_API
253 	if (strstr(device, "septel")) {
254 		return septel_open_live(device, snaplen, promisc, to_ms, ebuf);
255 	}
256 #endif /* HAVE_SEPTEL_API */
257 
258 	/* Allocate a handle for this session. */
259 
260 	handle = malloc(sizeof(*handle));
261 	if (handle == NULL) {
262 		snprintf(ebuf, PCAP_ERRBUF_SIZE, "malloc: %s",
263 			 pcap_strerror(errno));
264 		return NULL;
265 	}
266 
267 	/* Initialize some components of the pcap structure. */
268 
269 	memset(handle, 0, sizeof(*handle));
270 	handle->snapshot	= snaplen;
271 	handle->md.timeout	= to_ms;
272 
273 	/*
274 	 * NULL and "any" are special devices which give us the hint to
275 	 * monitor all devices.
276 	 */
277 	if (!device || strcmp(device, "any") == 0) {
278 		device			= NULL;
279 		handle->md.device	= strdup("any");
280 		if (promisc) {
281 			promisc = 0;
282 			/* Just a warning. */
283 			snprintf(ebuf, PCAP_ERRBUF_SIZE,
284 			    "Promiscuous mode not supported on the \"any\" device");
285 		}
286 
287 	} else
288 		handle->md.device	= strdup(device);
289 
290 	if (handle->md.device == NULL) {
291 		snprintf(ebuf, PCAP_ERRBUF_SIZE, "strdup: %s",
292 			 pcap_strerror(errno) );
293 		free(handle);
294 		return NULL;
295 	}
296 
297 	/*
298 	 * Current Linux kernels use the protocol family PF_PACKET to
299 	 * allow direct access to all packets on the network while
300 	 * older kernels had a special socket type SOCK_PACKET to
301 	 * implement this feature.
302 	 * While this old implementation is kind of obsolete we need
303 	 * to be compatible with older kernels for a while so we are
304 	 * trying both methods with the newer method preferred.
305 	 */
306 
307 	if ((err = live_open_new(handle, device, promisc, to_ms, ebuf)) == 1)
308 		live_open_ok = 1;
309 	else if (err == 0) {
310 		/* Non-fatal error; try old way */
311 		if (live_open_old(handle, device, promisc, to_ms, ebuf))
312 			live_open_ok = 1;
313 	}
314 	if (!live_open_ok) {
315 		/*
316 		 * Both methods to open the packet socket failed. Tidy
317 		 * up and report our failure (ebuf is expected to be
318 		 * set by the functions above).
319 		 */
320 
321 		if (handle->md.device != NULL)
322 			free(handle->md.device);
323 		free(handle);
324 		return NULL;
325 	}
326 
327 	/*
328 	 * Compute the buffer size.
329 	 *
330 	 * If we're using SOCK_PACKET, this might be a 2.0[.x] kernel,
331 	 * and might require special handling - check.
332 	 */
333 	if (handle->md.sock_packet && (uname(&utsname) < 0 ||
334 	    strncmp(utsname.release, "2.0", 3) == 0)) {
335 		/*
336 		 * We're using a SOCK_PACKET structure, and either
337 		 * we couldn't find out what kernel release this is,
338 		 * or it's a 2.0[.x] kernel.
339 		 *
340 		 * In the 2.0[.x] kernel, a "recvfrom()" on
341 		 * a SOCK_PACKET socket, with MSG_TRUNC set, will
342 		 * return the number of bytes read, so if we pass
343 		 * a length based on the snapshot length, it'll
344 		 * return the number of bytes from the packet
345 		 * copied to userland, not the actual length
346 		 * of the packet.
347 		 *
348 		 * This means that, for example, the IP dissector
349 		 * in tcpdump will get handed a packet length less
350 		 * than the length in the IP header, and will
351 		 * complain about "truncated-ip".
352 		 *
353 		 * So we don't bother trying to copy from the
354 		 * kernel only the bytes in which we're interested,
355 		 * but instead copy them all, just as the older
356 		 * versions of libpcap for Linux did.
357 		 *
358 		 * The buffer therefore needs to be big enough to
359 		 * hold the largest packet we can get from this
360 		 * device.  Unfortunately, we can't get the MRU
361 		 * of the network; we can only get the MTU.  The
362 		 * MTU may be too small, in which case a packet larger
363 		 * than the buffer size will be truncated *and* we
364 		 * won't get the actual packet size.
365 		 *
366 		 * However, if the snapshot length is larger than
367 		 * the buffer size based on the MTU, we use the
368 		 * snapshot length as the buffer size, instead;
369 		 * this means that with a sufficiently large snapshot
370 		 * length we won't artificially truncate packets
371 		 * to the MTU-based size.
372 		 *
373 		 * This mess just one of many problems with packet
374 		 * capture on 2.0[.x] kernels; you really want a
375 		 * 2.2[.x] or later kernel if you want packet capture
376 		 * to work well.
377 		 */
378 		mtu = iface_get_mtu(handle->fd, device, ebuf);
379 		if (mtu == -1) {
380 			pcap_close_linux(handle);
381 			free(handle);
382 			return NULL;
383 		}
384 		handle->bufsize = MAX_LINKHEADER_SIZE + mtu;
385 		if (handle->bufsize < handle->snapshot)
386 			handle->bufsize = handle->snapshot;
387 	} else {
388 		/*
389 		 * This is a 2.2[.x] or later kernel (we know that
390 		 * either because we're not using a SOCK_PACKET
391 		 * socket - PF_PACKET is supported only in 2.2
392 		 * and later kernels - or because we checked the
393 		 * kernel version).
394 		 *
395 		 * We can safely pass "recvfrom()" a byte count
396 		 * based on the snapshot length.
397 		 *
398 		 * If we're in cooked mode, make the snapshot length
399 		 * large enough to hold a "cooked mode" header plus
400 		 * 1 byte of packet data (so we don't pass a byte
401 		 * count of 0 to "recvfrom()").
402 		 */
403 		if (handle->md.cooked) {
404 			if (handle->snapshot < SLL_HDR_LEN + 1)
405 				handle->snapshot = SLL_HDR_LEN + 1;
406 		}
407 		handle->bufsize = handle->snapshot;
408 	}
409 
410 	/* Allocate the buffer */
411 
412 	handle->buffer	 = malloc(handle->bufsize + handle->offset);
413 	if (!handle->buffer) {
414 	        snprintf(ebuf, PCAP_ERRBUF_SIZE,
415 			 "malloc: %s", pcap_strerror(errno));
416 		pcap_close_linux(handle);
417 		free(handle);
418 		return NULL;
419 	}
420 
421 	/*
422 	 * "handle->fd" is a socket, so "select()" and "poll()"
423 	 * should work on it.
424 	 */
425 	handle->selectable_fd = handle->fd;
426 
427 	handle->read_op = pcap_read_linux;
428 	handle->inject_op = pcap_inject_linux;
429 	handle->setfilter_op = pcap_setfilter_linux;
430 	handle->setdirection_op = pcap_setdirection_linux;
431 	handle->set_datalink_op = NULL;	/* can't change data link type */
432 	handle->getnonblock_op = pcap_getnonblock_fd;
433 	handle->setnonblock_op = pcap_setnonblock_fd;
434 	handle->stats_op = pcap_stats_linux;
435 	handle->close_op = pcap_close_linux;
436 
437 	return handle;
438 }
439 
440 /*
441  *  Read at most max_packets from the capture stream and call the callback
442  *  for each of them. Returns the number of packets handled or -1 if an
443  *  error occured.
444  */
445 static int
446 pcap_read_linux(pcap_t *handle, int max_packets, pcap_handler callback, u_char *user)
447 {
448 	/*
449 	 * Currently, on Linux only one packet is delivered per read,
450 	 * so we don't loop.
451 	 */
452 	return pcap_read_packet(handle, callback, user);
453 }
454 
455 /*
456  *  Read a packet from the socket calling the handler provided by
457  *  the user. Returns the number of packets received or -1 if an
458  *  error occured.
459  */
460 static int
461 pcap_read_packet(pcap_t *handle, pcap_handler callback, u_char *userdata)
462 {
463 	u_char			*bp;
464 	int			offset;
465 #ifdef HAVE_PF_PACKET_SOCKETS
466 	struct sockaddr_ll	from;
467 	struct sll_header	*hdrp;
468 #else
469 	struct sockaddr		from;
470 #endif
471 	socklen_t		fromlen;
472 	int			packet_len, caplen;
473 	struct pcap_pkthdr	pcap_header;
474 
475 #ifdef HAVE_PF_PACKET_SOCKETS
476 	/*
477 	 * If this is a cooked device, leave extra room for a
478 	 * fake packet header.
479 	 */
480 	if (handle->md.cooked)
481 		offset = SLL_HDR_LEN;
482 	else
483 		offset = 0;
484 #else
485 	/*
486 	 * This system doesn't have PF_PACKET sockets, so it doesn't
487 	 * support cooked devices.
488 	 */
489 	offset = 0;
490 #endif
491 
492 	/* Receive a single packet from the kernel */
493 
494 	bp = handle->buffer + handle->offset;
495 	do {
496 		/*
497 		 * Has "pcap_breakloop()" been called?
498 		 */
499 		if (handle->break_loop) {
500 			/*
501 			 * Yes - clear the flag that indicates that it
502 			 * has, and return -2 as an indication that we
503 			 * were told to break out of the loop.
504 			 */
505 			handle->break_loop = 0;
506 			return -2;
507 		}
508 		fromlen = sizeof(from);
509 		packet_len = recvfrom(
510 			handle->fd, bp + offset,
511 			handle->bufsize - offset, MSG_TRUNC,
512 			(struct sockaddr *) &from, &fromlen);
513 	} while (packet_len == -1 && errno == EINTR);
514 
515 	/* Check if an error occured */
516 
517 	if (packet_len == -1) {
518 		if (errno == EAGAIN)
519 			return 0;	/* no packet there */
520 		else {
521 			snprintf(handle->errbuf, sizeof(handle->errbuf),
522 				 "recvfrom: %s", pcap_strerror(errno));
523 			return -1;
524 		}
525 	}
526 
527 #ifdef HAVE_PF_PACKET_SOCKETS
528 	if (!handle->md.sock_packet) {
529 		/*
530 		 * Unfortunately, there is a window between socket() and
531 		 * bind() where the kernel may queue packets from any
532 		 * interface.  If we're bound to a particular interface,
533 		 * discard packets not from that interface.
534 		 *
535 		 * (If socket filters are supported, we could do the
536 		 * same thing we do when changing the filter; however,
537 		 * that won't handle packet sockets without socket
538 		 * filter support, and it's a bit more complicated.
539 		 * It would save some instructions per packet, however.)
540 		 */
541 		if (handle->md.ifindex != -1 &&
542 		    from.sll_ifindex != handle->md.ifindex)
543 			return 0;
544 
545 		/*
546 		 * Do checks based on packet direction.
547 		 * We can only do this if we're using PF_PACKET; the
548 		 * address returned for SOCK_PACKET is a "sockaddr_pkt"
549 		 * which lacks the relevant packet type information.
550 		 */
551 		if (from.sll_pkttype == PACKET_OUTGOING) {
552 			/*
553 			 * Outgoing packet.
554 			 * If this is from the loopback device, reject it;
555 			 * we'll see the packet as an incoming packet as well,
556 			 * and we don't want to see it twice.
557 			 */
558 			if (from.sll_ifindex == handle->md.lo_ifindex)
559 				return 0;
560 
561 			/*
562 			 * If the user only wants incoming packets, reject it.
563 			 */
564 			if (handle->direction == PCAP_D_IN)
565 				return 0;
566 		} else {
567 			/*
568 			 * Incoming packet.
569 			 * If the user only wants outgoing packets, reject it.
570 			 */
571 			if (handle->direction == PCAP_D_OUT)
572 				return 0;
573 		}
574 	}
575 #endif
576 
577 #ifdef HAVE_PF_PACKET_SOCKETS
578 	/*
579 	 * If this is a cooked device, fill in the fake packet header.
580 	 */
581 	if (handle->md.cooked) {
582 		/*
583 		 * Add the length of the fake header to the length
584 		 * of packet data we read.
585 		 */
586 		packet_len += SLL_HDR_LEN;
587 
588 		hdrp = (struct sll_header *)bp;
589 
590 		/*
591 		 * Map the PACKET_ value to a LINUX_SLL_ value; we
592 		 * want the same numerical value to be used in
593 		 * the link-layer header even if the numerical values
594 		 * for the PACKET_ #defines change, so that programs
595 		 * that look at the packet type field will always be
596 		 * able to handle DLT_LINUX_SLL captures.
597 		 */
598 		switch (from.sll_pkttype) {
599 
600 		case PACKET_HOST:
601 			hdrp->sll_pkttype = htons(LINUX_SLL_HOST);
602 			break;
603 
604 		case PACKET_BROADCAST:
605 			hdrp->sll_pkttype = htons(LINUX_SLL_BROADCAST);
606 			break;
607 
608 		case PACKET_MULTICAST:
609 			hdrp->sll_pkttype = htons(LINUX_SLL_MULTICAST);
610 			break;
611 
612 		case PACKET_OTHERHOST:
613 			hdrp->sll_pkttype = htons(LINUX_SLL_OTHERHOST);
614 			break;
615 
616 		case PACKET_OUTGOING:
617 			hdrp->sll_pkttype = htons(LINUX_SLL_OUTGOING);
618 			break;
619 
620 		default:
621 			hdrp->sll_pkttype = -1;
622 			break;
623 		}
624 
625 		hdrp->sll_hatype = htons(from.sll_hatype);
626 		hdrp->sll_halen = htons(from.sll_halen);
627 		memcpy(hdrp->sll_addr, from.sll_addr,
628 		    (from.sll_halen > SLL_ADDRLEN) ?
629 		      SLL_ADDRLEN :
630 		      from.sll_halen);
631 		hdrp->sll_protocol = from.sll_protocol;
632 	}
633 #endif
634 
635 	/*
636 	 * XXX: According to the kernel source we should get the real
637 	 * packet len if calling recvfrom with MSG_TRUNC set. It does
638 	 * not seem to work here :(, but it is supported by this code
639 	 * anyway.
640 	 * To be honest the code RELIES on that feature so this is really
641 	 * broken with 2.2.x kernels.
642 	 * I spend a day to figure out what's going on and I found out
643 	 * that the following is happening:
644 	 *
645 	 * The packet comes from a random interface and the packet_rcv
646 	 * hook is called with a clone of the packet. That code inserts
647 	 * the packet into the receive queue of the packet socket.
648 	 * If a filter is attached to that socket that filter is run
649 	 * first - and there lies the problem. The default filter always
650 	 * cuts the packet at the snaplen:
651 	 *
652 	 * # tcpdump -d
653 	 * (000) ret      #68
654 	 *
655 	 * So the packet filter cuts down the packet. The recvfrom call
656 	 * says "hey, it's only 68 bytes, it fits into the buffer" with
657 	 * the result that we don't get the real packet length. This
658 	 * is valid at least until kernel 2.2.17pre6.
659 	 *
660 	 * We currently handle this by making a copy of the filter
661 	 * program, fixing all "ret" instructions with non-zero
662 	 * operands to have an operand of 65535 so that the filter
663 	 * doesn't truncate the packet, and supplying that modified
664 	 * filter to the kernel.
665 	 */
666 
667 	caplen = packet_len;
668 	if (caplen > handle->snapshot)
669 		caplen = handle->snapshot;
670 
671 	/* Run the packet filter if not using kernel filter */
672 	if (!handle->md.use_bpf && handle->fcode.bf_insns) {
673 		if (bpf_filter(handle->fcode.bf_insns, bp,
674 		                packet_len, caplen) == 0)
675 		{
676 			/* rejected by filter */
677 			return 0;
678 		}
679 	}
680 
681 	/* Fill in our own header data */
682 
683 	if (ioctl(handle->fd, SIOCGSTAMP, &pcap_header.ts) == -1) {
684 		snprintf(handle->errbuf, sizeof(handle->errbuf),
685 			 "SIOCGSTAMP: %s", pcap_strerror(errno));
686 		return -1;
687 	}
688 	pcap_header.caplen	= caplen;
689 	pcap_header.len		= packet_len;
690 
691 	/*
692 	 * Count the packet.
693 	 *
694 	 * Arguably, we should count them before we check the filter,
695 	 * as on many other platforms "ps_recv" counts packets
696 	 * handed to the filter rather than packets that passed
697 	 * the filter, but if filtering is done in the kernel, we
698 	 * can't get a count of packets that passed the filter,
699 	 * and that would mean the meaning of "ps_recv" wouldn't
700 	 * be the same on all Linux systems.
701 	 *
702 	 * XXX - it's not the same on all systems in any case;
703 	 * ideally, we should have a "get the statistics" call
704 	 * that supplies more counts and indicates which of them
705 	 * it supplies, so that we supply a count of packets
706 	 * handed to the filter only on platforms where that
707 	 * information is available.
708 	 *
709 	 * We count them here even if we can get the packet count
710 	 * from the kernel, as we can only determine at run time
711 	 * whether we'll be able to get it from the kernel (if
712 	 * HAVE_TPACKET_STATS isn't defined, we can't get it from
713 	 * the kernel, but if it is defined, the library might
714 	 * have been built with a 2.4 or later kernel, but we
715 	 * might be running on a 2.2[.x] kernel without Alexey
716 	 * Kuznetzov's turbopacket patches, and thus the kernel
717 	 * might not be able to supply those statistics).  We
718 	 * could, I guess, try, when opening the socket, to get
719 	 * the statistics, and if we can not increment the count
720 	 * here, but it's not clear that always incrementing
721 	 * the count is more expensive than always testing a flag
722 	 * in memory.
723 	 *
724 	 * We keep the count in "md.packets_read", and use that for
725 	 * "ps_recv" if we can't get the statistics from the kernel.
726 	 * We do that because, if we *can* get the statistics from
727 	 * the kernel, we use "md.stat.ps_recv" and "md.stat.ps_drop"
728 	 * as running counts, as reading the statistics from the
729 	 * kernel resets the kernel statistics, and if we directly
730 	 * increment "md.stat.ps_recv" here, that means it will
731 	 * count packets *twice* on systems where we can get kernel
732 	 * statistics - once here, and once in pcap_stats_linux().
733 	 */
734 	handle->md.packets_read++;
735 
736 	/* Call the user supplied callback function */
737 	callback(userdata, &pcap_header, bp);
738 
739 	return 1;
740 }
741 
742 static int
743 pcap_inject_linux(pcap_t *handle, const void *buf, size_t size)
744 {
745 	int ret;
746 
747 #ifdef HAVE_PF_PACKET_SOCKETS
748 	if (!handle->md.sock_packet) {
749 		/* PF_PACKET socket */
750 		if (handle->md.ifindex == -1) {
751 			/*
752 			 * We don't support sending on the "any" device.
753 			 */
754 			strlcpy(handle->errbuf,
755 			    "Sending packets isn't supported on the \"any\" device",
756 			    PCAP_ERRBUF_SIZE);
757 			return (-1);
758 		}
759 
760 		if (handle->md.cooked) {
761 			/*
762 			 * We don't support sending on the "any" device.
763 			 *
764 			 * XXX - how do you send on a bound cooked-mode
765 			 * socket?
766 			 * Is a "sendto()" required there?
767 			 */
768 			strlcpy(handle->errbuf,
769 			    "Sending packets isn't supported in cooked mode",
770 			    PCAP_ERRBUF_SIZE);
771 			return (-1);
772 		}
773 	}
774 #endif
775 
776 	ret = send(handle->fd, buf, size, 0);
777 	if (ret == -1) {
778 		snprintf(handle->errbuf, PCAP_ERRBUF_SIZE, "send: %s",
779 		    pcap_strerror(errno));
780 		return (-1);
781 	}
782 	return (ret);
783 }
784 
785 /*
786  *  Get the statistics for the given packet capture handle.
787  *  Reports the number of dropped packets iff the kernel supports
788  *  the PACKET_STATISTICS "getsockopt()" argument (2.4 and later
789  *  kernels, and 2.2[.x] kernels with Alexey Kuznetzov's turbopacket
790  *  patches); otherwise, that information isn't available, and we lie
791  *  and report 0 as the count of dropped packets.
792  */
793 static int
794 pcap_stats_linux(pcap_t *handle, struct pcap_stat *stats)
795 {
796 #ifdef HAVE_TPACKET_STATS
797 	struct tpacket_stats kstats;
798 	socklen_t len = sizeof (struct tpacket_stats);
799 #endif
800 
801 #ifdef HAVE_TPACKET_STATS
802 	/*
803 	 * Try to get the packet counts from the kernel.
804 	 */
805 	if (getsockopt(handle->fd, SOL_PACKET, PACKET_STATISTICS,
806 			&kstats, &len) > -1) {
807 		/*
808 		 * On systems where the PACKET_STATISTICS "getsockopt()"
809 		 * argument is supported on PF_PACKET sockets:
810 		 *
811 		 *	"ps_recv" counts only packets that *passed* the
812 		 *	filter, not packets that didn't pass the filter.
813 		 *	This includes packets later dropped because we
814 		 *	ran out of buffer space.
815 		 *
816 		 *	"ps_drop" counts packets dropped because we ran
817 		 *	out of buffer space.  It doesn't count packets
818 		 *	dropped by the interface driver.  It counts only
819 		 *	packets that passed the filter.
820 		 *
821 		 *	Both statistics include packets not yet read from
822 		 *	the kernel by libpcap, and thus not yet seen by
823 		 *	the application.
824 		 *
825 		 * In "linux/net/packet/af_packet.c", at least in the
826 		 * 2.4.9 kernel, "tp_packets" is incremented for every
827 		 * packet that passes the packet filter *and* is
828 		 * successfully queued on the socket; "tp_drops" is
829 		 * incremented for every packet dropped because there's
830 		 * not enough free space in the socket buffer.
831 		 *
832 		 * When the statistics are returned for a PACKET_STATISTICS
833 		 * "getsockopt()" call, "tp_drops" is added to "tp_packets",
834 		 * so that "tp_packets" counts all packets handed to
835 		 * the PF_PACKET socket, including packets dropped because
836 		 * there wasn't room on the socket buffer - but not
837 		 * including packets that didn't pass the filter.
838 		 *
839 		 * In the BSD BPF, the count of received packets is
840 		 * incremented for every packet handed to BPF, regardless
841 		 * of whether it passed the filter.
842 		 *
843 		 * We can't make "pcap_stats()" work the same on both
844 		 * platforms, but the best approximation is to return
845 		 * "tp_packets" as the count of packets and "tp_drops"
846 		 * as the count of drops.
847 		 *
848 		 * Keep a running total because each call to
849 		 *    getsockopt(handle->fd, SOL_PACKET, PACKET_STATISTICS, ....
850 		 * resets the counters to zero.
851 		 */
852 		handle->md.stat.ps_recv += kstats.tp_packets;
853 		handle->md.stat.ps_drop += kstats.tp_drops;
854 		*stats = handle->md.stat;
855 		return 0;
856 	}
857 	else
858 	{
859 		/*
860 		 * If the error was EOPNOTSUPP, fall through, so that
861 		 * if you build the library on a system with
862 		 * "struct tpacket_stats" and run it on a system
863 		 * that doesn't, it works as it does if the library
864 		 * is built on a system without "struct tpacket_stats".
865 		 */
866 		if (errno != EOPNOTSUPP) {
867 			snprintf(handle->errbuf, PCAP_ERRBUF_SIZE,
868 			    "pcap_stats: %s", pcap_strerror(errno));
869 			return -1;
870 		}
871 	}
872 #endif
873 	/*
874 	 * On systems where the PACKET_STATISTICS "getsockopt()" argument
875 	 * is not supported on PF_PACKET sockets:
876 	 *
877 	 *	"ps_recv" counts only packets that *passed* the filter,
878 	 *	not packets that didn't pass the filter.  It does not
879 	 *	count packets dropped because we ran out of buffer
880 	 *	space.
881 	 *
882 	 *	"ps_drop" is not supported.
883 	 *
884 	 *	"ps_recv" doesn't include packets not yet read from
885 	 *	the kernel by libpcap.
886 	 *
887 	 * We maintain the count of packets processed by libpcap in
888 	 * "md.packets_read", for reasons described in the comment
889 	 * at the end of pcap_read_packet().  We have no idea how many
890 	 * packets were dropped.
891 	 */
892 	stats->ps_recv = handle->md.packets_read;
893 	stats->ps_drop = 0;
894 	return 0;
895 }
896 
897 /*
898  * Description string for the "any" device.
899  */
900 static const char any_descr[] = "Pseudo-device that captures on all interfaces";
901 
902 int
903 pcap_platform_finddevs(pcap_if_t **alldevsp, char *errbuf)
904 {
905 	if (pcap_add_if(alldevsp, "any", 0, any_descr, errbuf) < 0)
906 		return (-1);
907 
908 #ifdef HAVE_DAG_API
909 	if (dag_platform_finddevs(alldevsp, errbuf) < 0)
910 		return (-1);
911 #endif /* HAVE_DAG_API */
912 
913 #ifdef HAVE_SEPTEL_API
914 	if (septel_platform_finddevs(alldevsp, errbuf) < 0)
915 		return (-1);
916 #endif /* HAVE_SEPTEL_API */
917 
918 	return (0);
919 }
920 
921 /*
922  *  Attach the given BPF code to the packet capture device.
923  */
924 static int
925 pcap_setfilter_linux(pcap_t *handle, struct bpf_program *filter)
926 {
927 #ifdef SO_ATTACH_FILTER
928 	struct sock_fprog	fcode;
929 	int			can_filter_in_kernel;
930 	int			err = 0;
931 #endif
932 
933 	if (!handle)
934 		return -1;
935 	if (!filter) {
936 	        strncpy(handle->errbuf, "setfilter: No filter specified",
937 			sizeof(handle->errbuf));
938 		return -1;
939 	}
940 
941 	/* Make our private copy of the filter */
942 
943 	if (install_bpf_program(handle, filter) < 0)
944 		/* install_bpf_program() filled in errbuf */
945 		return -1;
946 
947 	/*
948 	 * Run user level packet filter by default. Will be overriden if
949 	 * installing a kernel filter succeeds.
950 	 */
951 	handle->md.use_bpf = 0;
952 
953 	/* Install kernel level filter if possible */
954 
955 #ifdef SO_ATTACH_FILTER
956 #ifdef USHRT_MAX
957 	if (handle->fcode.bf_len > USHRT_MAX) {
958 		/*
959 		 * fcode.len is an unsigned short for current kernel.
960 		 * I have yet to see BPF-Code with that much
961 		 * instructions but still it is possible. So for the
962 		 * sake of correctness I added this check.
963 		 */
964 		fprintf(stderr, "Warning: Filter too complex for kernel\n");
965 		fcode.len = 0;
966 		fcode.filter = NULL;
967 		can_filter_in_kernel = 0;
968 	} else
969 #endif /* USHRT_MAX */
970 	{
971 		/*
972 		 * Oh joy, the Linux kernel uses struct sock_fprog instead
973 		 * of struct bpf_program and of course the length field is
974 		 * of different size. Pointed out by Sebastian
975 		 *
976 		 * Oh, and we also need to fix it up so that all "ret"
977 		 * instructions with non-zero operands have 65535 as the
978 		 * operand, and so that, if we're in cooked mode, all
979 		 * memory-reference instructions use special magic offsets
980 		 * in references to the link-layer header and assume that
981 		 * the link-layer payload begins at 0; "fix_program()"
982 		 * will do that.
983 		 */
984 		switch (fix_program(handle, &fcode)) {
985 
986 		case -1:
987 		default:
988 			/*
989 			 * Fatal error; just quit.
990 			 * (The "default" case shouldn't happen; we
991 			 * return -1 for that reason.)
992 			 */
993 			return -1;
994 
995 		case 0:
996 			/*
997 			 * The program performed checks that we can't make
998 			 * work in the kernel.
999 			 */
1000 			can_filter_in_kernel = 0;
1001 			break;
1002 
1003 		case 1:
1004 			/*
1005 			 * We have a filter that'll work in the kernel.
1006 			 */
1007 			can_filter_in_kernel = 1;
1008 			break;
1009 		}
1010 	}
1011 
1012 	if (can_filter_in_kernel) {
1013 		if ((err = set_kernel_filter(handle, &fcode)) == 0)
1014 		{
1015 			/* Installation succeded - using kernel filter. */
1016 			handle->md.use_bpf = 1;
1017 		}
1018 		else if (err == -1)	/* Non-fatal error */
1019 		{
1020 			/*
1021 			 * Print a warning if we weren't able to install
1022 			 * the filter for a reason other than "this kernel
1023 			 * isn't configured to support socket filters.
1024 			 */
1025 			if (errno != ENOPROTOOPT && errno != EOPNOTSUPP) {
1026 				fprintf(stderr,
1027 				    "Warning: Kernel filter failed: %s\n",
1028 					pcap_strerror(errno));
1029 			}
1030 		}
1031 	}
1032 
1033 	/*
1034 	 * If we're not using the kernel filter, get rid of any kernel
1035 	 * filter that might've been there before, e.g. because the
1036 	 * previous filter could work in the kernel, or because some other
1037 	 * code attached a filter to the socket by some means other than
1038 	 * calling "pcap_setfilter()".  Otherwise, the kernel filter may
1039 	 * filter out packets that would pass the new userland filter.
1040 	 */
1041 	if (!handle->md.use_bpf)
1042 		reset_kernel_filter(handle);
1043 
1044 	/*
1045 	 * Free up the copy of the filter that was made by "fix_program()".
1046 	 */
1047 	if (fcode.filter != NULL)
1048 		free(fcode.filter);
1049 
1050 	if (err == -2)
1051 		/* Fatal error */
1052 		return -1;
1053 #endif /* SO_ATTACH_FILTER */
1054 
1055 	return 0;
1056 }
1057 
1058 /*
1059  * Set direction flag: Which packets do we accept on a forwarding
1060  * single device? IN, OUT or both?
1061  */
1062 static int
1063 pcap_setdirection_linux(pcap_t *handle, pcap_direction_t d)
1064 {
1065 #ifdef HAVE_PF_PACKET_SOCKETS
1066 	if (!handle->md.sock_packet) {
1067 		handle->direction = d;
1068 		return 0;
1069 	}
1070 #endif
1071 	/*
1072 	 * We're not using PF_PACKET sockets, so we can't determine
1073 	 * the direction of the packet.
1074 	 */
1075 	snprintf(handle->errbuf, sizeof(handle->errbuf),
1076 	    "Setting direction is not supported on SOCK_PACKET sockets");
1077 	return -1;
1078 }
1079 
1080 /*
1081  *  Linux uses the ARP hardware type to identify the type of an
1082  *  interface. pcap uses the DLT_xxx constants for this. This
1083  *  function takes a pointer to a "pcap_t", and an ARPHRD_xxx
1084  *  constant, as arguments, and sets "handle->linktype" to the
1085  *  appropriate DLT_XXX constant and sets "handle->offset" to
1086  *  the appropriate value (to make "handle->offset" plus link-layer
1087  *  header length be a multiple of 4, so that the link-layer payload
1088  *  will be aligned on a 4-byte boundary when capturing packets).
1089  *  (If the offset isn't set here, it'll be 0; add code as appropriate
1090  *  for cases where it shouldn't be 0.)
1091  *
1092  *  If "cooked_ok" is non-zero, we can use DLT_LINUX_SLL and capture
1093  *  in cooked mode; otherwise, we can't use cooked mode, so we have
1094  *  to pick some type that works in raw mode, or fail.
1095  *
1096  *  Sets the link type to -1 if unable to map the type.
1097  */
1098 static void map_arphrd_to_dlt(pcap_t *handle, int arptype, int cooked_ok)
1099 {
1100 	switch (arptype) {
1101 
1102 	case ARPHRD_ETHER:
1103 		/*
1104 		 * This is (presumably) a real Ethernet capture; give it a
1105 		 * link-layer-type list with DLT_EN10MB and DLT_DOCSIS, so
1106 		 * that an application can let you choose it, in case you're
1107 		 * capturing DOCSIS traffic that a Cisco Cable Modem
1108 		 * Termination System is putting out onto an Ethernet (it
1109 		 * doesn't put an Ethernet header onto the wire, it puts raw
1110 		 * DOCSIS frames out on the wire inside the low-level
1111 		 * Ethernet framing).
1112 		 *
1113 		 * XXX - are there any sorts of "fake Ethernet" that have
1114 		 * ARPHRD_ETHER but that *shouldn't offer DLT_DOCSIS as
1115 		 * a Cisco CMTS won't put traffic onto it or get traffic
1116 		 * bridged onto it?  ISDN is handled in "live_open_new()",
1117 		 * as we fall back on cooked mode there; are there any
1118 		 * others?
1119 		 */
1120 		handle->dlt_list = (u_int *) malloc(sizeof(u_int) * 2);
1121 		/*
1122 		 * If that fails, just leave the list empty.
1123 		 */
1124 		if (handle->dlt_list != NULL) {
1125 			handle->dlt_list[0] = DLT_EN10MB;
1126 			handle->dlt_list[1] = DLT_DOCSIS;
1127 			handle->dlt_count = 2;
1128 		}
1129 		/* FALLTHROUGH */
1130 
1131 	case ARPHRD_METRICOM:
1132 	case ARPHRD_LOOPBACK:
1133 		handle->linktype = DLT_EN10MB;
1134 		handle->offset = 2;
1135 		break;
1136 
1137 	case ARPHRD_EETHER:
1138 		handle->linktype = DLT_EN3MB;
1139 		break;
1140 
1141 	case ARPHRD_AX25:
1142 		handle->linktype = DLT_AX25;
1143 		break;
1144 
1145 	case ARPHRD_PRONET:
1146 		handle->linktype = DLT_PRONET;
1147 		break;
1148 
1149 	case ARPHRD_CHAOS:
1150 		handle->linktype = DLT_CHAOS;
1151 		break;
1152 
1153 #ifndef ARPHRD_IEEE802_TR
1154 #define ARPHRD_IEEE802_TR 800	/* From Linux 2.4 */
1155 #endif
1156 	case ARPHRD_IEEE802_TR:
1157 	case ARPHRD_IEEE802:
1158 		handle->linktype = DLT_IEEE802;
1159 		handle->offset = 2;
1160 		break;
1161 
1162 	case ARPHRD_ARCNET:
1163 		handle->linktype = DLT_ARCNET_LINUX;
1164 		break;
1165 
1166 #ifndef ARPHRD_FDDI	/* From Linux 2.2.13 */
1167 #define ARPHRD_FDDI	774
1168 #endif
1169 	case ARPHRD_FDDI:
1170 		handle->linktype = DLT_FDDI;
1171 		handle->offset = 3;
1172 		break;
1173 
1174 #ifndef ARPHRD_ATM  /* FIXME: How to #include this? */
1175 #define ARPHRD_ATM 19
1176 #endif
1177 	case ARPHRD_ATM:
1178 		/*
1179 		 * The Classical IP implementation in ATM for Linux
1180 		 * supports both what RFC 1483 calls "LLC Encapsulation",
1181 		 * in which each packet has an LLC header, possibly
1182 		 * with a SNAP header as well, prepended to it, and
1183 		 * what RFC 1483 calls "VC Based Multiplexing", in which
1184 		 * different virtual circuits carry different network
1185 		 * layer protocols, and no header is prepended to packets.
1186 		 *
1187 		 * They both have an ARPHRD_ type of ARPHRD_ATM, so
1188 		 * you can't use the ARPHRD_ type to find out whether
1189 		 * captured packets will have an LLC header, and,
1190 		 * while there's a socket ioctl to *set* the encapsulation
1191 		 * type, there's no ioctl to *get* the encapsulation type.
1192 		 *
1193 		 * This means that
1194 		 *
1195 		 *	programs that dissect Linux Classical IP frames
1196 		 *	would have to check for an LLC header and,
1197 		 *	depending on whether they see one or not, dissect
1198 		 *	the frame as LLC-encapsulated or as raw IP (I
1199 		 *	don't know whether there's any traffic other than
1200 		 *	IP that would show up on the socket, or whether
1201 		 *	there's any support for IPv6 in the Linux
1202 		 *	Classical IP code);
1203 		 *
1204 		 *	filter expressions would have to compile into
1205 		 *	code that checks for an LLC header and does
1206 		 *	the right thing.
1207 		 *
1208 		 * Both of those are a nuisance - and, at least on systems
1209 		 * that support PF_PACKET sockets, we don't have to put
1210 		 * up with those nuisances; instead, we can just capture
1211 		 * in cooked mode.  That's what we'll do, if we can.
1212 		 * Otherwise, we'll just fail.
1213 		 */
1214 		if (cooked_ok)
1215 			handle->linktype = DLT_LINUX_SLL;
1216 		else
1217 			handle->linktype = -1;
1218 		break;
1219 
1220 #ifndef ARPHRD_IEEE80211  /* From Linux 2.4.6 */
1221 #define ARPHRD_IEEE80211 801
1222 #endif
1223 	case ARPHRD_IEEE80211:
1224 		handle->linktype = DLT_IEEE802_11;
1225 		break;
1226 
1227 #ifndef ARPHRD_IEEE80211_PRISM  /* From Linux 2.4.18 */
1228 #define ARPHRD_IEEE80211_PRISM 802
1229 #endif
1230 	case ARPHRD_IEEE80211_PRISM:
1231 		handle->linktype = DLT_PRISM_HEADER;
1232 		break;
1233 
1234 #ifndef ARPHRD_IEEE80211_RADIOTAP /* new */
1235 #define ARPHRD_IEEE80211_RADIOTAP 803
1236 #endif
1237 	case ARPHRD_IEEE80211_RADIOTAP:
1238 		handle->linktype = DLT_IEEE802_11_RADIO;
1239 		break;
1240 
1241 	case ARPHRD_PPP:
1242 		/*
1243 		 * Some PPP code in the kernel supplies no link-layer
1244 		 * header whatsoever to PF_PACKET sockets; other PPP
1245 		 * code supplies PPP link-layer headers ("syncppp.c");
1246 		 * some PPP code might supply random link-layer
1247 		 * headers (PPP over ISDN - there's code in Ethereal,
1248 		 * for example, to cope with PPP-over-ISDN captures
1249 		 * with which the Ethereal developers have had to cope,
1250 		 * heuristically trying to determine which of the
1251 		 * oddball link-layer headers particular packets have).
1252 		 *
1253 		 * As such, we just punt, and run all PPP interfaces
1254 		 * in cooked mode, if we can; otherwise, we just treat
1255 		 * it as DLT_RAW, for now - if somebody needs to capture,
1256 		 * on a 2.0[.x] kernel, on PPP devices that supply a
1257 		 * link-layer header, they'll have to add code here to
1258 		 * map to the appropriate DLT_ type (possibly adding a
1259 		 * new DLT_ type, if necessary).
1260 		 */
1261 		if (cooked_ok)
1262 			handle->linktype = DLT_LINUX_SLL;
1263 		else {
1264 			/*
1265 			 * XXX - handle ISDN types here?  We can't fall
1266 			 * back on cooked sockets, so we'd have to
1267 			 * figure out from the device name what type of
1268 			 * link-layer encapsulation it's using, and map
1269 			 * that to an appropriate DLT_ value, meaning
1270 			 * we'd map "isdnN" devices to DLT_RAW (they
1271 			 * supply raw IP packets with no link-layer
1272 			 * header) and "isdY" devices to a new DLT_I4L_IP
1273 			 * type that has only an Ethernet packet type as
1274 			 * a link-layer header.
1275 			 *
1276 			 * But sometimes we seem to get random crap
1277 			 * in the link-layer header when capturing on
1278 			 * ISDN devices....
1279 			 */
1280 			handle->linktype = DLT_RAW;
1281 		}
1282 		break;
1283 
1284 #ifndef ARPHRD_CISCO
1285 #define ARPHRD_CISCO 513 /* previously ARPHRD_HDLC */
1286 #endif
1287 	case ARPHRD_CISCO:
1288 		handle->linktype = DLT_C_HDLC;
1289 		break;
1290 
1291 	/* Not sure if this is correct for all tunnels, but it
1292 	 * works for CIPE */
1293 	case ARPHRD_TUNNEL:
1294 #ifndef ARPHRD_SIT
1295 #define ARPHRD_SIT 776	/* From Linux 2.2.13 */
1296 #endif
1297 	case ARPHRD_SIT:
1298 	case ARPHRD_CSLIP:
1299 	case ARPHRD_SLIP6:
1300 	case ARPHRD_CSLIP6:
1301 	case ARPHRD_ADAPT:
1302 	case ARPHRD_SLIP:
1303 #ifndef ARPHRD_RAWHDLC
1304 #define ARPHRD_RAWHDLC 518
1305 #endif
1306 	case ARPHRD_RAWHDLC:
1307 #ifndef ARPHRD_DLCI
1308 #define ARPHRD_DLCI 15
1309 #endif
1310 	case ARPHRD_DLCI:
1311 		/*
1312 		 * XXX - should some of those be mapped to DLT_LINUX_SLL
1313 		 * instead?  Should we just map all of them to DLT_LINUX_SLL?
1314 		 */
1315 		handle->linktype = DLT_RAW;
1316 		break;
1317 
1318 #ifndef ARPHRD_FRAD
1319 #define ARPHRD_FRAD 770
1320 #endif
1321 	case ARPHRD_FRAD:
1322 		handle->linktype = DLT_FRELAY;
1323 		break;
1324 
1325 	case ARPHRD_LOCALTLK:
1326 		handle->linktype = DLT_LTALK;
1327 		break;
1328 
1329 #ifndef ARPHRD_FCPP
1330 #define ARPHRD_FCPP	784
1331 #endif
1332 	case ARPHRD_FCPP:
1333 #ifndef ARPHRD_FCAL
1334 #define ARPHRD_FCAL	785
1335 #endif
1336 	case ARPHRD_FCAL:
1337 #ifndef ARPHRD_FCPL
1338 #define ARPHRD_FCPL	786
1339 #endif
1340 	case ARPHRD_FCPL:
1341 #ifndef ARPHRD_FCFABRIC
1342 #define ARPHRD_FCFABRIC	787
1343 #endif
1344 	case ARPHRD_FCFABRIC:
1345 		/*
1346 		 * We assume that those all mean RFC 2625 IP-over-
1347 		 * Fibre Channel, with the RFC 2625 header at
1348 		 * the beginning of the packet.
1349 		 */
1350 		handle->linktype = DLT_IP_OVER_FC;
1351 		break;
1352 
1353 #ifndef ARPHRD_IRDA
1354 #define ARPHRD_IRDA	783
1355 #endif
1356 	case ARPHRD_IRDA:
1357 		/* Don't expect IP packet out of this interfaces... */
1358 		handle->linktype = DLT_LINUX_IRDA;
1359 		/* We need to save packet direction for IrDA decoding,
1360 		 * so let's use "Linux-cooked" mode. Jean II */
1361 		//handle->md.cooked = 1;
1362 		break;
1363 
1364 	/* ARPHRD_LAPD is unofficial and randomly allocated, if reallocation
1365 	 * is needed, please report it to <daniele@orlandi.com> */
1366 #ifndef ARPHRD_LAPD
1367 #define ARPHRD_LAPD	8445
1368 #endif
1369 	case ARPHRD_LAPD:
1370 		/* Don't expect IP packet out of this interfaces... */
1371 		handle->linktype = DLT_LINUX_LAPD;
1372 		break;
1373 
1374 	default:
1375 		handle->linktype = -1;
1376 		break;
1377 	}
1378 }
1379 
1380 /* ===== Functions to interface to the newer kernels ================== */
1381 
1382 /*
1383  *  Try to open a packet socket using the new kernel interface.
1384  *  Returns 0 on failure.
1385  *  FIXME: 0 uses to mean success (Sebastian)
1386  */
1387 static int
1388 live_open_new(pcap_t *handle, const char *device, int promisc,
1389 	      int to_ms, char *ebuf)
1390 {
1391 #ifdef HAVE_PF_PACKET_SOCKETS
1392 	int			sock_fd = -1, arptype;
1393 	int			err;
1394 	int			fatal_err = 0;
1395 	struct packet_mreq	mr;
1396 
1397 	/* One shot loop used for error handling - bail out with break */
1398 
1399 	do {
1400 		/*
1401 		 * Open a socket with protocol family packet. If a device is
1402 		 * given we try to open it in raw mode otherwise we use
1403 		 * the cooked interface.
1404 		 */
1405 		sock_fd = device ?
1406 			socket(PF_PACKET, SOCK_RAW, htons(ETH_P_ALL))
1407 		      : socket(PF_PACKET, SOCK_DGRAM, htons(ETH_P_ALL));
1408 
1409 		if (sock_fd == -1) {
1410 			snprintf(ebuf, PCAP_ERRBUF_SIZE, "socket: %s",
1411 				 pcap_strerror(errno) );
1412 			break;
1413 		}
1414 
1415 		/* It seems the kernel supports the new interface. */
1416 		handle->md.sock_packet = 0;
1417 
1418 		/*
1419 		 * Get the interface index of the loopback device.
1420 		 * If the attempt fails, don't fail, just set the
1421 		 * "md.lo_ifindex" to -1.
1422 		 *
1423 		 * XXX - can there be more than one device that loops
1424 		 * packets back, i.e. devices other than "lo"?  If so,
1425 		 * we'd need to find them all, and have an array of
1426 		 * indices for them, and check all of them in
1427 		 * "pcap_read_packet()".
1428 		 */
1429 		handle->md.lo_ifindex = iface_get_id(sock_fd, "lo", ebuf);
1430 
1431 		/*
1432 		 * Default value for offset to align link-layer payload
1433 		 * on a 4-byte boundary.
1434 		 */
1435 		handle->offset	 = 0;
1436 
1437 		/*
1438 		 * What kind of frames do we have to deal with? Fall back
1439 		 * to cooked mode if we have an unknown interface type.
1440 		 */
1441 
1442 		if (device) {
1443 			/* Assume for now we don't need cooked mode. */
1444 			handle->md.cooked = 0;
1445 
1446 			arptype	= iface_get_arptype(sock_fd, device, ebuf);
1447 			if (arptype == -1) {
1448 				fatal_err = 1;
1449 				break;
1450 			}
1451 			map_arphrd_to_dlt(handle, arptype, 1);
1452 			if (handle->linktype == -1 ||
1453 			    handle->linktype == DLT_LINUX_SLL ||
1454 			    handle->linktype == DLT_LINUX_IRDA ||
1455 			    handle->linktype == DLT_LINUX_LAPD ||
1456 			    (handle->linktype == DLT_EN10MB &&
1457 			     (strncmp("isdn", device, 4) == 0 ||
1458 			      strncmp("isdY", device, 4) == 0))) {
1459 				/*
1460 				 * Unknown interface type (-1), or a
1461 				 * device we explicitly chose to run
1462 				 * in cooked mode (e.g., PPP devices),
1463 				 * or an ISDN device (whose link-layer
1464 				 * type we can only determine by using
1465 				 * APIs that may be different on different
1466 				 * kernels) - reopen in cooked mode.
1467 				 */
1468 				if (close(sock_fd) == -1) {
1469 					snprintf(ebuf, PCAP_ERRBUF_SIZE,
1470 						 "close: %s", pcap_strerror(errno));
1471 					break;
1472 				}
1473 				sock_fd = socket(PF_PACKET, SOCK_DGRAM,
1474 						 htons(ETH_P_ALL));
1475 				if (sock_fd == -1) {
1476 					snprintf(ebuf, PCAP_ERRBUF_SIZE,
1477 						 "socket: %s", pcap_strerror(errno));
1478 					break;
1479 				}
1480 				handle->md.cooked = 1;
1481 
1482 				/*
1483 				 * Get rid of any link-layer type list
1484 				 * we allocated - this only supports cooked
1485 				 * capture.
1486 				 */
1487 				if (handle->dlt_list != NULL) {
1488 					free(handle->dlt_list);
1489 					handle->dlt_list = NULL;
1490 					handle->dlt_count = 0;
1491 				}
1492 
1493 				if (handle->linktype == -1) {
1494 					/*
1495 					 * Warn that we're falling back on
1496 					 * cooked mode; we may want to
1497 					 * update "map_arphrd_to_dlt()"
1498 					 * to handle the new type.
1499 					 */
1500 					snprintf(ebuf, PCAP_ERRBUF_SIZE,
1501 						"arptype %d not "
1502 						"supported by libpcap - "
1503 						"falling back to cooked "
1504 						"socket",
1505 						arptype);
1506 				}
1507 				/* IrDA capture is not a real "cooked" capture,
1508 				 * it's IrLAP frames, not IP packets. */
1509 				if (handle->linktype != DLT_LINUX_IRDA &&
1510 			    		handle->linktype != DLT_LINUX_LAPD)
1511 					handle->linktype = DLT_LINUX_SLL;
1512 			}
1513 
1514 			handle->md.ifindex = iface_get_id(sock_fd, device, ebuf);
1515 			if (handle->md.ifindex == -1)
1516 				break;
1517 
1518 			if ((err = iface_bind(sock_fd, handle->md.ifindex,
1519 			    ebuf)) < 0) {
1520 				if (err == -2)
1521 					fatal_err = 1;
1522 				break;
1523 			}
1524 		} else {
1525 			/*
1526 			 * This is cooked mode.
1527 			 */
1528 			handle->md.cooked = 1;
1529 			handle->linktype = DLT_LINUX_SLL;
1530 
1531 			/*
1532 			 * We're not bound to a device.
1533 			 * XXX - true?  Or true only if we're using
1534 			 * the "any" device?
1535 			 * For now, we're using this as an indication
1536 			 * that we can't transmit; stop doing that only
1537 			 * if we figure out how to transmit in cooked
1538 			 * mode.
1539 			 */
1540 			handle->md.ifindex = -1;
1541 		}
1542 
1543 		/*
1544 		 * Select promiscuous mode on if "promisc" is set.
1545 		 *
1546 		 * Do not turn allmulti mode on if we don't select
1547 		 * promiscuous mode - on some devices (e.g., Orinoco
1548 		 * wireless interfaces), allmulti mode isn't supported
1549 		 * and the driver implements it by turning promiscuous
1550 		 * mode on, and that screws up the operation of the
1551 		 * card as a normal networking interface, and on no
1552 		 * other platform I know of does starting a non-
1553 		 * promiscuous capture affect which multicast packets
1554 		 * are received by the interface.
1555 		 */
1556 
1557 		/*
1558 		 * Hmm, how can we set promiscuous mode on all interfaces?
1559 		 * I am not sure if that is possible at all.
1560 		 */
1561 
1562 		if (device && promisc) {
1563 			memset(&mr, 0, sizeof(mr));
1564 			mr.mr_ifindex = handle->md.ifindex;
1565 			mr.mr_type    = PACKET_MR_PROMISC;
1566 			if (setsockopt(sock_fd, SOL_PACKET,
1567 				PACKET_ADD_MEMBERSHIP, &mr, sizeof(mr)) == -1)
1568 			{
1569 				snprintf(ebuf, PCAP_ERRBUF_SIZE,
1570 					"setsockopt: %s", pcap_strerror(errno));
1571 				break;
1572 			}
1573 		}
1574 
1575 		/* Save the socket FD in the pcap structure */
1576 
1577 		handle->fd 	 = sock_fd;
1578 
1579 		return 1;
1580 
1581 	} while(0);
1582 
1583 	if (sock_fd != -1)
1584 		close(sock_fd);
1585 
1586 	if (fatal_err) {
1587 		/*
1588 		 * Get rid of any link-layer type list we allocated.
1589 		 */
1590 		if (handle->dlt_list != NULL)
1591 			free(handle->dlt_list);
1592 		return -2;
1593 	} else
1594 		return 0;
1595 #else
1596 	strncpy(ebuf,
1597 		"New packet capturing interface not supported by build "
1598 		"environment", PCAP_ERRBUF_SIZE);
1599 	return 0;
1600 #endif
1601 }
1602 
1603 #ifdef HAVE_PF_PACKET_SOCKETS
1604 /*
1605  *  Return the index of the given device name. Fill ebuf and return
1606  *  -1 on failure.
1607  */
1608 static int
1609 iface_get_id(int fd, const char *device, char *ebuf)
1610 {
1611 	struct ifreq	ifr;
1612 
1613 	memset(&ifr, 0, sizeof(ifr));
1614 	strncpy(ifr.ifr_name, device, sizeof(ifr.ifr_name));
1615 
1616 	if (ioctl(fd, SIOCGIFINDEX, &ifr) == -1) {
1617 		snprintf(ebuf, PCAP_ERRBUF_SIZE,
1618 			 "SIOCGIFINDEX: %s", pcap_strerror(errno));
1619 		return -1;
1620 	}
1621 
1622 	return ifr.ifr_ifindex;
1623 }
1624 
1625 /*
1626  *  Bind the socket associated with FD to the given device.
1627  */
1628 static int
1629 iface_bind(int fd, int ifindex, char *ebuf)
1630 {
1631 	struct sockaddr_ll	sll;
1632 	int			err;
1633 	socklen_t		errlen = sizeof(err);
1634 
1635 	memset(&sll, 0, sizeof(sll));
1636 	sll.sll_family		= AF_PACKET;
1637 	sll.sll_ifindex		= ifindex;
1638 	sll.sll_protocol	= htons(ETH_P_ALL);
1639 
1640 	if (bind(fd, (struct sockaddr *) &sll, sizeof(sll)) == -1) {
1641 		snprintf(ebuf, PCAP_ERRBUF_SIZE,
1642 			 "bind: %s", pcap_strerror(errno));
1643 		return -1;
1644 	}
1645 
1646 	/* Any pending errors, e.g., network is down? */
1647 
1648 	if (getsockopt(fd, SOL_SOCKET, SO_ERROR, &err, &errlen) == -1) {
1649 		snprintf(ebuf, PCAP_ERRBUF_SIZE,
1650 			"getsockopt: %s", pcap_strerror(errno));
1651 		return -2;
1652 	}
1653 
1654 	if (err > 0) {
1655 		snprintf(ebuf, PCAP_ERRBUF_SIZE,
1656 			"bind: %s", pcap_strerror(err));
1657 		return -2;
1658 	}
1659 
1660 	return 0;
1661 }
1662 
1663 #endif
1664 
1665 
1666 /* ===== Functions to interface to the older kernels ================== */
1667 
1668 /*
1669  * With older kernels promiscuous mode is kind of interesting because we
1670  * have to reset the interface before exiting. The problem can't really
1671  * be solved without some daemon taking care of managing usage counts.
1672  * If we put the interface into promiscuous mode, we set a flag indicating
1673  * that we must take it out of that mode when the interface is closed,
1674  * and, when closing the interface, if that flag is set we take it out
1675  * of promiscuous mode.
1676  */
1677 
1678 /*
1679  * List of pcaps for which we turned promiscuous mode on by hand.
1680  * If there are any such pcaps, we arrange to call "pcap_close_all()"
1681  * when we exit, and have it close all of them to turn promiscuous mode
1682  * off.
1683  */
1684 static struct pcap *pcaps_to_close;
1685 
1686 /*
1687  * TRUE if we've already called "atexit()" to cause "pcap_close_all()" to
1688  * be called on exit.
1689  */
1690 static int did_atexit;
1691 
1692 static void	pcap_close_all(void)
1693 {
1694 	struct pcap *handle;
1695 
1696 	while ((handle = pcaps_to_close) != NULL)
1697 		pcap_close(handle);
1698 }
1699 
1700 static void	pcap_close_linux( pcap_t *handle )
1701 {
1702 	struct pcap	*p, *prevp;
1703 	struct ifreq	ifr;
1704 
1705 	if (handle->md.clear_promisc) {
1706 		/*
1707 		 * We put the interface into promiscuous mode; take
1708 		 * it out of promiscuous mode.
1709 		 *
1710 		 * XXX - if somebody else wants it in promiscuous mode,
1711 		 * this code cannot know that, so it'll take it out
1712 		 * of promiscuous mode.  That's not fixable in 2.0[.x]
1713 		 * kernels.
1714 		 */
1715 		memset(&ifr, 0, sizeof(ifr));
1716 		strncpy(ifr.ifr_name, handle->md.device, sizeof(ifr.ifr_name));
1717 		if (ioctl(handle->fd, SIOCGIFFLAGS, &ifr) == -1) {
1718 			fprintf(stderr,
1719 			    "Can't restore interface flags (SIOCGIFFLAGS failed: %s).\n"
1720 			    "Please adjust manually.\n"
1721 			    "Hint: This can't happen with Linux >= 2.2.0.\n",
1722 			    strerror(errno));
1723 		} else {
1724 			if (ifr.ifr_flags & IFF_PROMISC) {
1725 				/*
1726 				 * Promiscuous mode is currently on; turn it
1727 				 * off.
1728 				 */
1729 				ifr.ifr_flags &= ~IFF_PROMISC;
1730 				if (ioctl(handle->fd, SIOCSIFFLAGS, &ifr) == -1) {
1731 					fprintf(stderr,
1732 					    "Can't restore interface flags (SIOCSIFFLAGS failed: %s).\n"
1733 					    "Please adjust manually.\n"
1734 					    "Hint: This can't happen with Linux >= 2.2.0.\n",
1735 					    strerror(errno));
1736 				}
1737 			}
1738 		}
1739 
1740 		/*
1741 		 * Take this pcap out of the list of pcaps for which we
1742 		 * have to take the interface out of promiscuous mode.
1743 		 */
1744 		for (p = pcaps_to_close, prevp = NULL; p != NULL;
1745 		    prevp = p, p = p->md.next) {
1746 			if (p == handle) {
1747 				/*
1748 				 * Found it.  Remove it from the list.
1749 				 */
1750 				if (prevp == NULL) {
1751 					/*
1752 					 * It was at the head of the list.
1753 					 */
1754 					pcaps_to_close = p->md.next;
1755 				} else {
1756 					/*
1757 					 * It was in the middle of the list.
1758 					 */
1759 					prevp->md.next = p->md.next;
1760 				}
1761 				break;
1762 			}
1763 		}
1764 	}
1765 
1766 	if (handle->md.device != NULL)
1767 		free(handle->md.device);
1768 	handle->md.device = NULL;
1769 	pcap_close_common(handle);
1770 }
1771 
1772 /*
1773  *  Try to open a packet socket using the old kernel interface.
1774  *  Returns 0 on failure.
1775  *  FIXME: 0 uses to mean success (Sebastian)
1776  */
1777 static int
1778 live_open_old(pcap_t *handle, const char *device, int promisc,
1779 	      int to_ms, char *ebuf)
1780 {
1781 	int		arptype;
1782 	struct ifreq	ifr;
1783 
1784 	do {
1785 		/* Open the socket */
1786 
1787 		handle->fd = socket(PF_INET, SOCK_PACKET, htons(ETH_P_ALL));
1788 		if (handle->fd == -1) {
1789 			snprintf(ebuf, PCAP_ERRBUF_SIZE,
1790 				 "socket: %s", pcap_strerror(errno));
1791 			break;
1792 		}
1793 
1794 		/* It worked - we are using the old interface */
1795 		handle->md.sock_packet = 1;
1796 
1797 		/* ...which means we get the link-layer header. */
1798 		handle->md.cooked = 0;
1799 
1800 		/* Bind to the given device */
1801 
1802 		if (!device) {
1803 		        strncpy(ebuf, "pcap_open_live: The \"any\" device isn't supported on 2.0[.x]-kernel systems",
1804 				PCAP_ERRBUF_SIZE);
1805 			break;
1806 		}
1807 		if (iface_bind_old(handle->fd, device, ebuf) == -1)
1808 			break;
1809 
1810 		/*
1811 		 * Try to get the link-layer type.
1812 		 */
1813 		arptype = iface_get_arptype(handle->fd, device, ebuf);
1814 		if (arptype == -1)
1815 			break;
1816 
1817 		/*
1818 		 * Try to find the DLT_ type corresponding to that
1819 		 * link-layer type.
1820 		 */
1821 		map_arphrd_to_dlt(handle, arptype, 0);
1822 		if (handle->linktype == -1) {
1823 			snprintf(ebuf, PCAP_ERRBUF_SIZE,
1824 				 "unknown arptype %d", arptype);
1825 			break;
1826 		}
1827 
1828 		/* Go to promisc mode if requested */
1829 
1830 		if (promisc) {
1831 			memset(&ifr, 0, sizeof(ifr));
1832 			strncpy(ifr.ifr_name, device, sizeof(ifr.ifr_name));
1833 			if (ioctl(handle->fd, SIOCGIFFLAGS, &ifr) == -1) {
1834 				snprintf(ebuf, PCAP_ERRBUF_SIZE,
1835 					 "SIOCGIFFLAGS: %s", pcap_strerror(errno));
1836 				break;
1837 			}
1838 			if ((ifr.ifr_flags & IFF_PROMISC) == 0) {
1839 				/*
1840 				 * Promiscuous mode isn't currently on,
1841 				 * so turn it on, and remember that
1842 				 * we should turn it off when the
1843 				 * pcap_t is closed.
1844 				 */
1845 
1846 				/*
1847 				 * If we haven't already done so, arrange
1848 				 * to have "pcap_close_all()" called when
1849 				 * we exit.
1850 				 */
1851 				if (!did_atexit) {
1852 					if (atexit(pcap_close_all) == -1) {
1853 						/*
1854 						 * "atexit()" failed; don't
1855 						 * put the interface in
1856 						 * promiscuous mode, just
1857 						 * give up.
1858 						 */
1859 						strncpy(ebuf, "atexit failed",
1860 							PCAP_ERRBUF_SIZE);
1861 						break;
1862 					}
1863 					did_atexit = 1;
1864 				}
1865 
1866 				ifr.ifr_flags |= IFF_PROMISC;
1867 				if (ioctl(handle->fd, SIOCSIFFLAGS, &ifr) == -1) {
1868 				        snprintf(ebuf, PCAP_ERRBUF_SIZE,
1869 						 "SIOCSIFFLAGS: %s",
1870 						 pcap_strerror(errno));
1871 					break;
1872 				}
1873 				handle->md.clear_promisc = 1;
1874 
1875 				/*
1876 				 * Add this to the list of pcaps
1877 				 * to close when we exit.
1878 				 */
1879 				handle->md.next = pcaps_to_close;
1880 				pcaps_to_close = handle;
1881 			}
1882 		}
1883 
1884 		/*
1885 		 * Default value for offset to align link-layer payload
1886 		 * on a 4-byte boundary.
1887 		 */
1888 		handle->offset	 = 0;
1889 
1890 		return 1;
1891 
1892 	} while (0);
1893 
1894 	pcap_close_linux(handle);
1895 	return 0;
1896 }
1897 
1898 /*
1899  *  Bind the socket associated with FD to the given device using the
1900  *  interface of the old kernels.
1901  */
1902 static int
1903 iface_bind_old(int fd, const char *device, char *ebuf)
1904 {
1905 	struct sockaddr	saddr;
1906 	int		err;
1907 	socklen_t	errlen = sizeof(err);
1908 
1909 	memset(&saddr, 0, sizeof(saddr));
1910 	strncpy(saddr.sa_data, device, sizeof(saddr.sa_data));
1911 	if (bind(fd, &saddr, sizeof(saddr)) == -1) {
1912 		snprintf(ebuf, PCAP_ERRBUF_SIZE,
1913 			 "bind: %s", pcap_strerror(errno));
1914 		return -1;
1915 	}
1916 
1917 	/* Any pending errors, e.g., network is down? */
1918 
1919 	if (getsockopt(fd, SOL_SOCKET, SO_ERROR, &err, &errlen) == -1) {
1920 		snprintf(ebuf, PCAP_ERRBUF_SIZE,
1921 			"getsockopt: %s", pcap_strerror(errno));
1922 		return -1;
1923 	}
1924 
1925 	if (err > 0) {
1926 		snprintf(ebuf, PCAP_ERRBUF_SIZE,
1927 			"bind: %s", pcap_strerror(err));
1928 		return -1;
1929 	}
1930 
1931 	return 0;
1932 }
1933 
1934 
1935 /* ===== System calls available on all supported kernels ============== */
1936 
1937 /*
1938  *  Query the kernel for the MTU of the given interface.
1939  */
1940 static int
1941 iface_get_mtu(int fd, const char *device, char *ebuf)
1942 {
1943 	struct ifreq	ifr;
1944 
1945 	if (!device)
1946 		return BIGGER_THAN_ALL_MTUS;
1947 
1948 	memset(&ifr, 0, sizeof(ifr));
1949 	strncpy(ifr.ifr_name, device, sizeof(ifr.ifr_name));
1950 
1951 	if (ioctl(fd, SIOCGIFMTU, &ifr) == -1) {
1952 		snprintf(ebuf, PCAP_ERRBUF_SIZE,
1953 			 "SIOCGIFMTU: %s", pcap_strerror(errno));
1954 		return -1;
1955 	}
1956 
1957 	return ifr.ifr_mtu;
1958 }
1959 
1960 /*
1961  *  Get the hardware type of the given interface as ARPHRD_xxx constant.
1962  */
1963 static int
1964 iface_get_arptype(int fd, const char *device, char *ebuf)
1965 {
1966 	struct ifreq	ifr;
1967 
1968 	memset(&ifr, 0, sizeof(ifr));
1969 	strncpy(ifr.ifr_name, device, sizeof(ifr.ifr_name));
1970 
1971 	if (ioctl(fd, SIOCGIFHWADDR, &ifr) == -1) {
1972 		snprintf(ebuf, PCAP_ERRBUF_SIZE,
1973 			 "SIOCGIFHWADDR: %s", pcap_strerror(errno));
1974 		return -1;
1975 	}
1976 
1977 	return ifr.ifr_hwaddr.sa_family;
1978 }
1979 
1980 #ifdef SO_ATTACH_FILTER
1981 static int
1982 fix_program(pcap_t *handle, struct sock_fprog *fcode)
1983 {
1984 	size_t prog_size;
1985 	register int i;
1986 	register struct bpf_insn *p;
1987 	struct bpf_insn *f;
1988 	int len;
1989 
1990 	/*
1991 	 * Make a copy of the filter, and modify that copy if
1992 	 * necessary.
1993 	 */
1994 	prog_size = sizeof(*handle->fcode.bf_insns) * handle->fcode.bf_len;
1995 	len = handle->fcode.bf_len;
1996 	f = (struct bpf_insn *)malloc(prog_size);
1997 	if (f == NULL) {
1998 		snprintf(handle->errbuf, sizeof(handle->errbuf),
1999 			 "malloc: %s", pcap_strerror(errno));
2000 		return -1;
2001 	}
2002 	memcpy(f, handle->fcode.bf_insns, prog_size);
2003 	fcode->len = len;
2004 	fcode->filter = (struct sock_filter *) f;
2005 
2006 	for (i = 0; i < len; ++i) {
2007 		p = &f[i];
2008 		/*
2009 		 * What type of instruction is this?
2010 		 */
2011 		switch (BPF_CLASS(p->code)) {
2012 
2013 		case BPF_RET:
2014 			/*
2015 			 * It's a return instruction; is the snapshot
2016 			 * length a constant, rather than the contents
2017 			 * of the accumulator?
2018 			 */
2019 			if (BPF_MODE(p->code) == BPF_K) {
2020 				/*
2021 				 * Yes - if the value to be returned,
2022 				 * i.e. the snapshot length, is anything
2023 				 * other than 0, make it 65535, so that
2024 				 * the packet is truncated by "recvfrom()",
2025 				 * not by the filter.
2026 				 *
2027 				 * XXX - there's nothing we can easily do
2028 				 * if it's getting the value from the
2029 				 * accumulator; we'd have to insert
2030 				 * code to force non-zero values to be
2031 				 * 65535.
2032 				 */
2033 				if (p->k != 0)
2034 					p->k = 65535;
2035 			}
2036 			break;
2037 
2038 		case BPF_LD:
2039 		case BPF_LDX:
2040 			/*
2041 			 * It's a load instruction; is it loading
2042 			 * from the packet?
2043 			 */
2044 			switch (BPF_MODE(p->code)) {
2045 
2046 			case BPF_ABS:
2047 			case BPF_IND:
2048 			case BPF_MSH:
2049 				/*
2050 				 * Yes; are we in cooked mode?
2051 				 */
2052 				if (handle->md.cooked) {
2053 					/*
2054 					 * Yes, so we need to fix this
2055 					 * instruction.
2056 					 */
2057 					if (fix_offset(p) < 0) {
2058 						/*
2059 						 * We failed to do so.
2060 						 * Return 0, so our caller
2061 						 * knows to punt to userland.
2062 						 */
2063 						return 0;
2064 					}
2065 				}
2066 				break;
2067 			}
2068 			break;
2069 		}
2070 	}
2071 	return 1;	/* we succeeded */
2072 }
2073 
2074 static int
2075 fix_offset(struct bpf_insn *p)
2076 {
2077 	/*
2078 	 * What's the offset?
2079 	 */
2080 	if (p->k >= SLL_HDR_LEN) {
2081 		/*
2082 		 * It's within the link-layer payload; that starts at an
2083 		 * offset of 0, as far as the kernel packet filter is
2084 		 * concerned, so subtract the length of the link-layer
2085 		 * header.
2086 		 */
2087 		p->k -= SLL_HDR_LEN;
2088 	} else if (p->k == 14) {
2089 		/*
2090 		 * It's the protocol field; map it to the special magic
2091 		 * kernel offset for that field.
2092 		 */
2093 		p->k = SKF_AD_OFF + SKF_AD_PROTOCOL;
2094 	} else {
2095 		/*
2096 		 * It's within the header, but it's not one of those
2097 		 * fields; we can't do that in the kernel, so punt
2098 		 * to userland.
2099 		 */
2100 		return -1;
2101 	}
2102 	return 0;
2103 }
2104 
2105 static int
2106 set_kernel_filter(pcap_t *handle, struct sock_fprog *fcode)
2107 {
2108 	int total_filter_on = 0;
2109 	int save_mode;
2110 	int ret;
2111 	int save_errno;
2112 
2113 	/*
2114 	 * The socket filter code doesn't discard all packets queued
2115 	 * up on the socket when the filter is changed; this means
2116 	 * that packets that don't match the new filter may show up
2117 	 * after the new filter is put onto the socket, if those
2118 	 * packets haven't yet been read.
2119 	 *
2120 	 * This means, for example, that if you do a tcpdump capture
2121 	 * with a filter, the first few packets in the capture might
2122 	 * be packets that wouldn't have passed the filter.
2123 	 *
2124 	 * We therefore discard all packets queued up on the socket
2125 	 * when setting a kernel filter.  (This isn't an issue for
2126 	 * userland filters, as the userland filtering is done after
2127 	 * packets are queued up.)
2128 	 *
2129 	 * To flush those packets, we put the socket in read-only mode,
2130 	 * and read packets from the socket until there are no more to
2131 	 * read.
2132 	 *
2133 	 * In order to keep that from being an infinite loop - i.e.,
2134 	 * to keep more packets from arriving while we're draining
2135 	 * the queue - we put the "total filter", which is a filter
2136 	 * that rejects all packets, onto the socket before draining
2137 	 * the queue.
2138 	 *
2139 	 * This code deliberately ignores any errors, so that you may
2140 	 * get bogus packets if an error occurs, rather than having
2141 	 * the filtering done in userland even if it could have been
2142 	 * done in the kernel.
2143 	 */
2144 	if (setsockopt(handle->fd, SOL_SOCKET, SO_ATTACH_FILTER,
2145 		       &total_fcode, sizeof(total_fcode)) == 0) {
2146 		char drain[1];
2147 
2148 		/*
2149 		 * Note that we've put the total filter onto the socket.
2150 		 */
2151 		total_filter_on = 1;
2152 
2153 		/*
2154 		 * Save the socket's current mode, and put it in
2155 		 * non-blocking mode; we drain it by reading packets
2156 		 * until we get an error (which is normally a
2157 		 * "nothing more to be read" error).
2158 		 */
2159 		save_mode = fcntl(handle->fd, F_GETFL, 0);
2160 		if (save_mode != -1 &&
2161 		    fcntl(handle->fd, F_SETFL, save_mode | O_NONBLOCK) >= 0) {
2162 			while (recv(handle->fd, &drain, sizeof drain,
2163 			       MSG_TRUNC) >= 0)
2164 				;
2165 			save_errno = errno;
2166 			fcntl(handle->fd, F_SETFL, save_mode);
2167 			if (save_errno != EAGAIN) {
2168 				/* Fatal error */
2169 				reset_kernel_filter(handle);
2170 				snprintf(handle->errbuf, sizeof(handle->errbuf),
2171 				 "recv: %s", pcap_strerror(save_errno));
2172 				return -2;
2173 			}
2174 		}
2175 	}
2176 
2177 	/*
2178 	 * Now attach the new filter.
2179 	 */
2180 	ret = setsockopt(handle->fd, SOL_SOCKET, SO_ATTACH_FILTER,
2181 			 fcode, sizeof(*fcode));
2182 	if (ret == -1 && total_filter_on) {
2183 		/*
2184 		 * Well, we couldn't set that filter on the socket,
2185 		 * but we could set the total filter on the socket.
2186 		 *
2187 		 * This could, for example, mean that the filter was
2188 		 * too big to put into the kernel, so we'll have to
2189 		 * filter in userland; in any case, we'll be doing
2190 		 * filtering in userland, so we need to remove the
2191 		 * total filter so we see packets.
2192 		 */
2193 		save_errno = errno;
2194 
2195 		/*
2196 		 * XXX - if this fails, we're really screwed;
2197 		 * we have the total filter on the socket,
2198 		 * and it won't come off.  What do we do then?
2199 		 */
2200 		reset_kernel_filter(handle);
2201 
2202 		errno = save_errno;
2203 	}
2204 	return ret;
2205 }
2206 
2207 static int
2208 reset_kernel_filter(pcap_t *handle)
2209 {
2210 	/*
2211 	 * setsockopt() barfs unless it get a dummy parameter.
2212 	 * valgrind whines unless the value is initialized,
2213 	 * as it has no idea that setsockopt() ignores its
2214 	 * parameter.
2215 	 */
2216 	int dummy = 0;
2217 
2218 	return setsockopt(handle->fd, SOL_SOCKET, SO_DETACH_FILTER,
2219 				   &dummy, sizeof(dummy));
2220 }
2221 #endif
2222