xref: /freebsd/contrib/libpcap/pcap-linux.c (revision 282a3889ebf826db9839be296ff1dd903f6d6d6e)
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 1.110.2.6 2005/08/16 04:25:26 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 		handle->bufsize = handle->snapshot;
399 	}
400 
401 	/* Allocate the buffer */
402 
403 	handle->buffer	 = malloc(handle->bufsize + handle->offset);
404 	if (!handle->buffer) {
405 	        snprintf(ebuf, PCAP_ERRBUF_SIZE,
406 			 "malloc: %s", pcap_strerror(errno));
407 		pcap_close_linux(handle);
408 		free(handle);
409 		return NULL;
410 	}
411 
412 	/*
413 	 * "handle->fd" is a socket, so "select()" and "poll()"
414 	 * should work on it.
415 	 */
416 	handle->selectable_fd = handle->fd;
417 
418 	handle->read_op = pcap_read_linux;
419 	handle->inject_op = pcap_inject_linux;
420 	handle->setfilter_op = pcap_setfilter_linux;
421 	handle->setdirection_op = pcap_setdirection_linux;
422 	handle->set_datalink_op = NULL;	/* can't change data link type */
423 	handle->getnonblock_op = pcap_getnonblock_fd;
424 	handle->setnonblock_op = pcap_setnonblock_fd;
425 	handle->stats_op = pcap_stats_linux;
426 	handle->close_op = pcap_close_linux;
427 
428 	return handle;
429 }
430 
431 /*
432  *  Read at most max_packets from the capture stream and call the callback
433  *  for each of them. Returns the number of packets handled or -1 if an
434  *  error occured.
435  */
436 static int
437 pcap_read_linux(pcap_t *handle, int max_packets, pcap_handler callback, u_char *user)
438 {
439 	/*
440 	 * Currently, on Linux only one packet is delivered per read,
441 	 * so we don't loop.
442 	 */
443 	return pcap_read_packet(handle, callback, user);
444 }
445 
446 /*
447  *  Read a packet from the socket calling the handler provided by
448  *  the user. Returns the number of packets received or -1 if an
449  *  error occured.
450  */
451 static int
452 pcap_read_packet(pcap_t *handle, pcap_handler callback, u_char *userdata)
453 {
454 	u_char			*bp;
455 	int			offset;
456 #ifdef HAVE_PF_PACKET_SOCKETS
457 	struct sockaddr_ll	from;
458 	struct sll_header	*hdrp;
459 #else
460 	struct sockaddr		from;
461 #endif
462 	socklen_t		fromlen;
463 	int			packet_len, caplen;
464 	struct pcap_pkthdr	pcap_header;
465 
466 #ifdef HAVE_PF_PACKET_SOCKETS
467 	/*
468 	 * If this is a cooked device, leave extra room for a
469 	 * fake packet header.
470 	 */
471 	if (handle->md.cooked)
472 		offset = SLL_HDR_LEN;
473 	else
474 		offset = 0;
475 #else
476 	/*
477 	 * This system doesn't have PF_PACKET sockets, so it doesn't
478 	 * support cooked devices.
479 	 */
480 	offset = 0;
481 #endif
482 
483 	/* Receive a single packet from the kernel */
484 
485 	bp = handle->buffer + handle->offset;
486 	do {
487 		/*
488 		 * Has "pcap_breakloop()" been called?
489 		 */
490 		if (handle->break_loop) {
491 			/*
492 			 * Yes - clear the flag that indicates that it
493 			 * has, and return -2 as an indication that we
494 			 * were told to break out of the loop.
495 			 */
496 			handle->break_loop = 0;
497 			return -2;
498 		}
499 		fromlen = sizeof(from);
500 		packet_len = recvfrom(
501 			handle->fd, bp + offset,
502 			handle->bufsize - offset, MSG_TRUNC,
503 			(struct sockaddr *) &from, &fromlen);
504 	} while (packet_len == -1 && errno == EINTR);
505 
506 	/* Check if an error occured */
507 
508 	if (packet_len == -1) {
509 		if (errno == EAGAIN)
510 			return 0;	/* no packet there */
511 		else {
512 			snprintf(handle->errbuf, sizeof(handle->errbuf),
513 				 "recvfrom: %s", pcap_strerror(errno));
514 			return -1;
515 		}
516 	}
517 
518 #ifdef HAVE_PF_PACKET_SOCKETS
519 	if (!handle->md.sock_packet) {
520 		/*
521 		 * Do checks based on packet direction.
522 		 * We can only do this if we're using PF_PACKET; the
523 		 * address returned for SOCK_PACKET is a "sockaddr_pkt"
524 		 * which lacks the relevant packet type information.
525 		 */
526 		if (from.sll_pkttype == PACKET_OUTGOING) {
527 			/*
528 			 * Outgoing packet.
529 			 * If this is from the loopback device, reject it;
530 			 * we'll see the packet as an incoming packet as well,
531 			 * and we don't want to see it twice.
532 			 */
533 			if (from.sll_ifindex == handle->md.lo_ifindex)
534 				return 0;
535 
536 			/*
537 			 * If the user only wants incoming packets, reject it.
538 			 */
539 			if (handle->direction == PCAP_D_IN)
540 				return 0;
541 		} else {
542 			/*
543 			 * Incoming packet.
544 			 * If the user only wants outgoing packets, reject it.
545 			 */
546 			if (handle->direction == PCAP_D_OUT)
547 				return 0;
548 		}
549 	}
550 #endif
551 
552 #ifdef HAVE_PF_PACKET_SOCKETS
553 	/*
554 	 * If this is a cooked device, fill in the fake packet header.
555 	 */
556 	if (handle->md.cooked) {
557 		/*
558 		 * Add the length of the fake header to the length
559 		 * of packet data we read.
560 		 */
561 		packet_len += SLL_HDR_LEN;
562 
563 		hdrp = (struct sll_header *)bp;
564 
565 		/*
566 		 * Map the PACKET_ value to a LINUX_SLL_ value; we
567 		 * want the same numerical value to be used in
568 		 * the link-layer header even if the numerical values
569 		 * for the PACKET_ #defines change, so that programs
570 		 * that look at the packet type field will always be
571 		 * able to handle DLT_LINUX_SLL captures.
572 		 */
573 		switch (from.sll_pkttype) {
574 
575 		case PACKET_HOST:
576 			hdrp->sll_pkttype = htons(LINUX_SLL_HOST);
577 			break;
578 
579 		case PACKET_BROADCAST:
580 			hdrp->sll_pkttype = htons(LINUX_SLL_BROADCAST);
581 			break;
582 
583 		case PACKET_MULTICAST:
584 			hdrp->sll_pkttype = htons(LINUX_SLL_MULTICAST);
585 			break;
586 
587 		case PACKET_OTHERHOST:
588 			hdrp->sll_pkttype = htons(LINUX_SLL_OTHERHOST);
589 			break;
590 
591 		case PACKET_OUTGOING:
592 			hdrp->sll_pkttype = htons(LINUX_SLL_OUTGOING);
593 			break;
594 
595 		default:
596 			hdrp->sll_pkttype = -1;
597 			break;
598 		}
599 
600 		hdrp->sll_hatype = htons(from.sll_hatype);
601 		hdrp->sll_halen = htons(from.sll_halen);
602 		memcpy(hdrp->sll_addr, from.sll_addr,
603 		    (from.sll_halen > SLL_ADDRLEN) ?
604 		      SLL_ADDRLEN :
605 		      from.sll_halen);
606 		hdrp->sll_protocol = from.sll_protocol;
607 	}
608 #endif
609 
610 	/*
611 	 * XXX: According to the kernel source we should get the real
612 	 * packet len if calling recvfrom with MSG_TRUNC set. It does
613 	 * not seem to work here :(, but it is supported by this code
614 	 * anyway.
615 	 * To be honest the code RELIES on that feature so this is really
616 	 * broken with 2.2.x kernels.
617 	 * I spend a day to figure out what's going on and I found out
618 	 * that the following is happening:
619 	 *
620 	 * The packet comes from a random interface and the packet_rcv
621 	 * hook is called with a clone of the packet. That code inserts
622 	 * the packet into the receive queue of the packet socket.
623 	 * If a filter is attached to that socket that filter is run
624 	 * first - and there lies the problem. The default filter always
625 	 * cuts the packet at the snaplen:
626 	 *
627 	 * # tcpdump -d
628 	 * (000) ret      #68
629 	 *
630 	 * So the packet filter cuts down the packet. The recvfrom call
631 	 * says "hey, it's only 68 bytes, it fits into the buffer" with
632 	 * the result that we don't get the real packet length. This
633 	 * is valid at least until kernel 2.2.17pre6.
634 	 *
635 	 * We currently handle this by making a copy of the filter
636 	 * program, fixing all "ret" instructions with non-zero
637 	 * operands to have an operand of 65535 so that the filter
638 	 * doesn't truncate the packet, and supplying that modified
639 	 * filter to the kernel.
640 	 */
641 
642 	caplen = packet_len;
643 	if (caplen > handle->snapshot)
644 		caplen = handle->snapshot;
645 
646 	/* Run the packet filter if not using kernel filter */
647 	if (!handle->md.use_bpf && handle->fcode.bf_insns) {
648 		if (bpf_filter(handle->fcode.bf_insns, bp,
649 		                packet_len, caplen) == 0)
650 		{
651 			/* rejected by filter */
652 			return 0;
653 		}
654 	}
655 
656 	/* Fill in our own header data */
657 
658 	if (ioctl(handle->fd, SIOCGSTAMP, &pcap_header.ts) == -1) {
659 		snprintf(handle->errbuf, sizeof(handle->errbuf),
660 			 "ioctl: %s", pcap_strerror(errno));
661 		return -1;
662 	}
663 	pcap_header.caplen	= caplen;
664 	pcap_header.len		= packet_len;
665 
666 	/*
667 	 * Count the packet.
668 	 *
669 	 * Arguably, we should count them before we check the filter,
670 	 * as on many other platforms "ps_recv" counts packets
671 	 * handed to the filter rather than packets that passed
672 	 * the filter, but if filtering is done in the kernel, we
673 	 * can't get a count of packets that passed the filter,
674 	 * and that would mean the meaning of "ps_recv" wouldn't
675 	 * be the same on all Linux systems.
676 	 *
677 	 * XXX - it's not the same on all systems in any case;
678 	 * ideally, we should have a "get the statistics" call
679 	 * that supplies more counts and indicates which of them
680 	 * it supplies, so that we supply a count of packets
681 	 * handed to the filter only on platforms where that
682 	 * information is available.
683 	 *
684 	 * We count them here even if we can get the packet count
685 	 * from the kernel, as we can only determine at run time
686 	 * whether we'll be able to get it from the kernel (if
687 	 * HAVE_TPACKET_STATS isn't defined, we can't get it from
688 	 * the kernel, but if it is defined, the library might
689 	 * have been built with a 2.4 or later kernel, but we
690 	 * might be running on a 2.2[.x] kernel without Alexey
691 	 * Kuznetzov's turbopacket patches, and thus the kernel
692 	 * might not be able to supply those statistics).  We
693 	 * could, I guess, try, when opening the socket, to get
694 	 * the statistics, and if we can not increment the count
695 	 * here, but it's not clear that always incrementing
696 	 * the count is more expensive than always testing a flag
697 	 * in memory.
698 	 */
699 	handle->md.stat.ps_recv++;
700 
701 	/* Call the user supplied callback function */
702 	callback(userdata, &pcap_header, bp);
703 
704 	return 1;
705 }
706 
707 static int
708 pcap_inject_linux(pcap_t *handle, const void *buf, size_t size)
709 {
710 	int ret;
711 
712 #ifdef HAVE_PF_PACKET_SOCKETS
713 	if (!handle->md.sock_packet) {
714 		/* PF_PACKET socket */
715 		if (handle->md.ifindex == -1) {
716 			/*
717 			 * We don't support sending on the "any" device.
718 			 */
719 			strlcpy(handle->errbuf,
720 			    "Sending packets isn't supported on the \"any\" device",
721 			    PCAP_ERRBUF_SIZE);
722 			return (-1);
723 		}
724 
725 		if (handle->md.cooked) {
726 			/*
727 			 * We don't support sending on the "any" device.
728 			 *
729 			 * XXX - how do you send on a bound cooked-mode
730 			 * socket?
731 			 * Is a "sendto()" required there?
732 			 */
733 			strlcpy(handle->errbuf,
734 			    "Sending packets isn't supported in cooked mode",
735 			    PCAP_ERRBUF_SIZE);
736 			return (-1);
737 		}
738 	}
739 #endif
740 
741 	ret = send(handle->fd, buf, size, 0);
742 	if (ret == -1) {
743 		snprintf(handle->errbuf, PCAP_ERRBUF_SIZE, "send: %s",
744 		    pcap_strerror(errno));
745 		return (-1);
746 	}
747 	return (ret);
748 }
749 
750 /*
751  *  Get the statistics for the given packet capture handle.
752  *  Reports the number of dropped packets iff the kernel supports
753  *  the PACKET_STATISTICS "getsockopt()" argument (2.4 and later
754  *  kernels, and 2.2[.x] kernels with Alexey Kuznetzov's turbopacket
755  *  patches); otherwise, that information isn't available, and we lie
756  *  and report 0 as the count of dropped packets.
757  */
758 static int
759 pcap_stats_linux(pcap_t *handle, struct pcap_stat *stats)
760 {
761 #ifdef HAVE_TPACKET_STATS
762 	struct tpacket_stats kstats;
763 	socklen_t len = sizeof (struct tpacket_stats);
764 #endif
765 
766 #ifdef HAVE_TPACKET_STATS
767 	/*
768 	 * Try to get the packet counts from the kernel.
769 	 */
770 	if (getsockopt(handle->fd, SOL_PACKET, PACKET_STATISTICS,
771 			&kstats, &len) > -1) {
772 		/*
773 		 * In "linux/net/packet/af_packet.c", at least in the
774 		 * 2.4.9 kernel, "tp_packets" is incremented for every
775 		 * packet that passes the packet filter *and* is
776 		 * successfully queued on the socket; "tp_drops" is
777 		 * incremented for every packet dropped because there's
778 		 * not enough free space in the socket buffer.
779 		 *
780 		 * When the statistics are returned for a PACKET_STATISTICS
781 		 * "getsockopt()" call, "tp_drops" is added to "tp_packets",
782 		 * so that "tp_packets" counts all packets handed to
783 		 * the PF_PACKET socket, including packets dropped because
784 		 * there wasn't room on the socket buffer - but not
785 		 * including packets that didn't pass the filter.
786 		 *
787 		 * In the BSD BPF, the count of received packets is
788 		 * incremented for every packet handed to BPF, regardless
789 		 * of whether it passed the filter.
790 		 *
791 		 * We can't make "pcap_stats()" work the same on both
792 		 * platforms, but the best approximation is to return
793 		 * "tp_packets" as the count of packets and "tp_drops"
794 		 * as the count of drops.
795 		 *
796 		 * Keep a running total because each call to
797 		 *    getsockopt(handle->fd, SOL_PACKET, PACKET_STATISTICS, ....
798 		 * resets the counters to zero.
799 		 */
800 		handle->md.stat.ps_recv += kstats.tp_packets;
801 		handle->md.stat.ps_drop += kstats.tp_drops;
802 	}
803 	else
804 	{
805 		/*
806 		 * If the error was EOPNOTSUPP, fall through, so that
807 		 * if you build the library on a system with
808 		 * "struct tpacket_stats" and run it on a system
809 		 * that doesn't, it works as it does if the library
810 		 * is built on a system without "struct tpacket_stats".
811 		 */
812 		if (errno != EOPNOTSUPP) {
813 			snprintf(handle->errbuf, PCAP_ERRBUF_SIZE,
814 			    "pcap_stats: %s", pcap_strerror(errno));
815 			return -1;
816 		}
817 	}
818 #endif
819 	/*
820 	 * On systems where the PACKET_STATISTICS "getsockopt()" argument
821 	 * is supported on PF_PACKET sockets:
822 	 *
823 	 *	"ps_recv" counts only packets that *passed* the filter,
824 	 *	not packets that didn't pass the filter.  This includes
825 	 *	packets later dropped because we ran out of buffer space.
826 	 *
827 	 *	"ps_drop" counts packets dropped because we ran out of
828 	 *	buffer space.  It doesn't count packets dropped by the
829 	 *	interface driver.  It counts only packets that passed
830 	 *	the filter.
831 	 *
832 	 *	Both statistics include packets not yet read from the
833 	 *	kernel by libpcap, and thus not yet seen by the application.
834 	 *
835 	 * On systems where the PACKET_STATISTICS "getsockopt()" argument
836 	 * is not supported on PF_PACKET sockets:
837 	 *
838 	 *	"ps_recv" counts only packets that *passed* the filter,
839 	 *	not packets that didn't pass the filter.  It does not
840 	 *	count packets dropped because we ran out of buffer
841 	 *	space.
842 	 *
843 	 *	"ps_drop" is not supported.
844 	 *
845 	 *	"ps_recv" doesn't include packets not yet read from
846 	 *	the kernel by libpcap.
847 	 */
848 	*stats = handle->md.stat;
849 	return 0;
850 }
851 
852 /*
853  * Description string for the "any" device.
854  */
855 static const char any_descr[] = "Pseudo-device that captures on all interfaces";
856 
857 int
858 pcap_platform_finddevs(pcap_if_t **alldevsp, char *errbuf)
859 {
860 	if (pcap_add_if(alldevsp, "any", 0, any_descr, errbuf) < 0)
861 		return (-1);
862 
863 #ifdef HAVE_DAG_API
864 	if (dag_platform_finddevs(alldevsp, errbuf) < 0)
865 		return (-1);
866 #endif /* HAVE_DAG_API */
867 
868 #ifdef HAVE_SEPTEL_API
869 	if (septel_platform_finddevs(alldevsp, errbuf) < 0)
870 		return (-1);
871 #endif /* HAVE_SEPTEL_API */
872 
873 	return (0);
874 }
875 
876 /*
877  *  Attach the given BPF code to the packet capture device.
878  */
879 static int
880 pcap_setfilter_linux(pcap_t *handle, struct bpf_program *filter)
881 {
882 #ifdef SO_ATTACH_FILTER
883 	struct sock_fprog	fcode;
884 	int			can_filter_in_kernel;
885 	int			err = 0;
886 #endif
887 
888 	if (!handle)
889 		return -1;
890 	if (!filter) {
891 	        strncpy(handle->errbuf, "setfilter: No filter specified",
892 			sizeof(handle->errbuf));
893 		return -1;
894 	}
895 
896 	/* Make our private copy of the filter */
897 
898 	if (install_bpf_program(handle, filter) < 0)
899 		/* install_bpf_program() filled in errbuf */
900 		return -1;
901 
902 	/*
903 	 * Run user level packet filter by default. Will be overriden if
904 	 * installing a kernel filter succeeds.
905 	 */
906 	handle->md.use_bpf = 0;
907 
908 	/* Install kernel level filter if possible */
909 
910 #ifdef SO_ATTACH_FILTER
911 #ifdef USHRT_MAX
912 	if (handle->fcode.bf_len > USHRT_MAX) {
913 		/*
914 		 * fcode.len is an unsigned short for current kernel.
915 		 * I have yet to see BPF-Code with that much
916 		 * instructions but still it is possible. So for the
917 		 * sake of correctness I added this check.
918 		 */
919 		fprintf(stderr, "Warning: Filter too complex for kernel\n");
920 		fcode.filter = NULL;
921 		can_filter_in_kernel = 0;
922 	} else
923 #endif /* USHRT_MAX */
924 	{
925 		/*
926 		 * Oh joy, the Linux kernel uses struct sock_fprog instead
927 		 * of struct bpf_program and of course the length field is
928 		 * of different size. Pointed out by Sebastian
929 		 *
930 		 * Oh, and we also need to fix it up so that all "ret"
931 		 * instructions with non-zero operands have 65535 as the
932 		 * operand, and so that, if we're in cooked mode, all
933 		 * memory-reference instructions use special magic offsets
934 		 * in references to the link-layer header and assume that
935 		 * the link-layer payload begins at 0; "fix_program()"
936 		 * will do that.
937 		 */
938 		switch (fix_program(handle, &fcode)) {
939 
940 		case -1:
941 		default:
942 			/*
943 			 * Fatal error; just quit.
944 			 * (The "default" case shouldn't happen; we
945 			 * return -1 for that reason.)
946 			 */
947 			return -1;
948 
949 		case 0:
950 			/*
951 			 * The program performed checks that we can't make
952 			 * work in the kernel.
953 			 */
954 			can_filter_in_kernel = 0;
955 			break;
956 
957 		case 1:
958 			/*
959 			 * We have a filter that'll work in the kernel.
960 			 */
961 			can_filter_in_kernel = 1;
962 			break;
963 		}
964 	}
965 
966 	if (can_filter_in_kernel) {
967 		if ((err = set_kernel_filter(handle, &fcode)) == 0)
968 		{
969 			/* Installation succeded - using kernel filter. */
970 			handle->md.use_bpf = 1;
971 		}
972 		else if (err == -1)	/* Non-fatal error */
973 		{
974 			/*
975 			 * Print a warning if we weren't able to install
976 			 * the filter for a reason other than "this kernel
977 			 * isn't configured to support socket filters.
978 			 */
979 			if (errno != ENOPROTOOPT && errno != EOPNOTSUPP) {
980 				fprintf(stderr,
981 				    "Warning: Kernel filter failed: %s\n",
982 					pcap_strerror(errno));
983 			}
984 		}
985 	}
986 
987 	/*
988 	 * If we're not using the kernel filter, get rid of any kernel
989 	 * filter that might've been there before, e.g. because the
990 	 * previous filter could work in the kernel, or because some other
991 	 * code attached a filter to the socket by some means other than
992 	 * calling "pcap_setfilter()".  Otherwise, the kernel filter may
993 	 * filter out packets that would pass the new userland filter.
994 	 */
995 	if (!handle->md.use_bpf)
996 		reset_kernel_filter(handle);
997 
998 	/*
999 	 * Free up the copy of the filter that was made by "fix_program()".
1000 	 */
1001 	if (fcode.filter != NULL)
1002 		free(fcode.filter);
1003 
1004 	if (err == -2)
1005 		/* Fatal error */
1006 		return -1;
1007 #endif /* SO_ATTACH_FILTER */
1008 
1009 	return 0;
1010 }
1011 
1012 /*
1013  * Set direction flag: Which packets do we accept on a forwarding
1014  * single device? IN, OUT or both?
1015  */
1016 static int
1017 pcap_setdirection_linux(pcap_t *handle, pcap_direction_t d)
1018 {
1019 #ifdef HAVE_PF_PACKET_SOCKETS
1020 	if (!handle->md.sock_packet) {
1021 		handle->direction = d;
1022 		return 0;
1023 	}
1024 #endif
1025 	/*
1026 	 * We're not using PF_PACKET sockets, so we can't determine
1027 	 * the direction of the packet.
1028 	 */
1029 	snprintf(handle->errbuf, sizeof(handle->errbuf),
1030 	    "Setting direction is not supported on SOCK_PACKET sockets");
1031 	return -1;
1032 }
1033 
1034 /*
1035  *  Linux uses the ARP hardware type to identify the type of an
1036  *  interface. pcap uses the DLT_xxx constants for this. This
1037  *  function takes a pointer to a "pcap_t", and an ARPHRD_xxx
1038  *  constant, as arguments, and sets "handle->linktype" to the
1039  *  appropriate DLT_XXX constant and sets "handle->offset" to
1040  *  the appropriate value (to make "handle->offset" plus link-layer
1041  *  header length be a multiple of 4, so that the link-layer payload
1042  *  will be aligned on a 4-byte boundary when capturing packets).
1043  *  (If the offset isn't set here, it'll be 0; add code as appropriate
1044  *  for cases where it shouldn't be 0.)
1045  *
1046  *  If "cooked_ok" is non-zero, we can use DLT_LINUX_SLL and capture
1047  *  in cooked mode; otherwise, we can't use cooked mode, so we have
1048  *  to pick some type that works in raw mode, or fail.
1049  *
1050  *  Sets the link type to -1 if unable to map the type.
1051  */
1052 static void map_arphrd_to_dlt(pcap_t *handle, int arptype, int cooked_ok)
1053 {
1054 	switch (arptype) {
1055 
1056 	case ARPHRD_ETHER:
1057 		/*
1058 		 * This is (presumably) a real Ethernet capture; give it a
1059 		 * link-layer-type list with DLT_EN10MB and DLT_DOCSIS, so
1060 		 * that an application can let you choose it, in case you're
1061 		 * capturing DOCSIS traffic that a Cisco Cable Modem
1062 		 * Termination System is putting out onto an Ethernet (it
1063 		 * doesn't put an Ethernet header onto the wire, it puts raw
1064 		 * DOCSIS frames out on the wire inside the low-level
1065 		 * Ethernet framing).
1066 		 *
1067 		 * XXX - are there any sorts of "fake Ethernet" that have
1068 		 * ARPHRD_ETHER but that *shouldn't offer DLT_DOCSIS as
1069 		 * a Cisco CMTS won't put traffic onto it or get traffic
1070 		 * bridged onto it?  ISDN is handled in "live_open_new()",
1071 		 * as we fall back on cooked mode there; are there any
1072 		 * others?
1073 		 */
1074 		handle->dlt_list = (u_int *) malloc(sizeof(u_int) * 2);
1075 		/*
1076 		 * If that fails, just leave the list empty.
1077 		 */
1078 		if (handle->dlt_list != NULL) {
1079 			handle->dlt_list[0] = DLT_EN10MB;
1080 			handle->dlt_list[1] = DLT_DOCSIS;
1081 			handle->dlt_count = 2;
1082 		}
1083 		/* FALLTHROUGH */
1084 
1085 	case ARPHRD_METRICOM:
1086 	case ARPHRD_LOOPBACK:
1087 		handle->linktype = DLT_EN10MB;
1088 		handle->offset = 2;
1089 		break;
1090 
1091 	case ARPHRD_EETHER:
1092 		handle->linktype = DLT_EN3MB;
1093 		break;
1094 
1095 	case ARPHRD_AX25:
1096 		handle->linktype = DLT_AX25;
1097 		break;
1098 
1099 	case ARPHRD_PRONET:
1100 		handle->linktype = DLT_PRONET;
1101 		break;
1102 
1103 	case ARPHRD_CHAOS:
1104 		handle->linktype = DLT_CHAOS;
1105 		break;
1106 
1107 #ifndef ARPHRD_IEEE802_TR
1108 #define ARPHRD_IEEE802_TR 800	/* From Linux 2.4 */
1109 #endif
1110 	case ARPHRD_IEEE802_TR:
1111 	case ARPHRD_IEEE802:
1112 		handle->linktype = DLT_IEEE802;
1113 		handle->offset = 2;
1114 		break;
1115 
1116 	case ARPHRD_ARCNET:
1117 		handle->linktype = DLT_ARCNET_LINUX;
1118 		break;
1119 
1120 #ifndef ARPHRD_FDDI	/* From Linux 2.2.13 */
1121 #define ARPHRD_FDDI	774
1122 #endif
1123 	case ARPHRD_FDDI:
1124 		handle->linktype = DLT_FDDI;
1125 		handle->offset = 3;
1126 		break;
1127 
1128 #ifndef ARPHRD_ATM  /* FIXME: How to #include this? */
1129 #define ARPHRD_ATM 19
1130 #endif
1131 	case ARPHRD_ATM:
1132 		/*
1133 		 * The Classical IP implementation in ATM for Linux
1134 		 * supports both what RFC 1483 calls "LLC Encapsulation",
1135 		 * in which each packet has an LLC header, possibly
1136 		 * with a SNAP header as well, prepended to it, and
1137 		 * what RFC 1483 calls "VC Based Multiplexing", in which
1138 		 * different virtual circuits carry different network
1139 		 * layer protocols, and no header is prepended to packets.
1140 		 *
1141 		 * They both have an ARPHRD_ type of ARPHRD_ATM, so
1142 		 * you can't use the ARPHRD_ type to find out whether
1143 		 * captured packets will have an LLC header, and,
1144 		 * while there's a socket ioctl to *set* the encapsulation
1145 		 * type, there's no ioctl to *get* the encapsulation type.
1146 		 *
1147 		 * This means that
1148 		 *
1149 		 *	programs that dissect Linux Classical IP frames
1150 		 *	would have to check for an LLC header and,
1151 		 *	depending on whether they see one or not, dissect
1152 		 *	the frame as LLC-encapsulated or as raw IP (I
1153 		 *	don't know whether there's any traffic other than
1154 		 *	IP that would show up on the socket, or whether
1155 		 *	there's any support for IPv6 in the Linux
1156 		 *	Classical IP code);
1157 		 *
1158 		 *	filter expressions would have to compile into
1159 		 *	code that checks for an LLC header and does
1160 		 *	the right thing.
1161 		 *
1162 		 * Both of those are a nuisance - and, at least on systems
1163 		 * that support PF_PACKET sockets, we don't have to put
1164 		 * up with those nuisances; instead, we can just capture
1165 		 * in cooked mode.  That's what we'll do, if we can.
1166 		 * Otherwise, we'll just fail.
1167 		 */
1168 		if (cooked_ok)
1169 			handle->linktype = DLT_LINUX_SLL;
1170 		else
1171 			handle->linktype = -1;
1172 		break;
1173 
1174 #ifndef ARPHRD_IEEE80211  /* From Linux 2.4.6 */
1175 #define ARPHRD_IEEE80211 801
1176 #endif
1177 	case ARPHRD_IEEE80211:
1178 		handle->linktype = DLT_IEEE802_11;
1179 		break;
1180 
1181 #ifndef ARPHRD_IEEE80211_PRISM  /* From Linux 2.4.18 */
1182 #define ARPHRD_IEEE80211_PRISM 802
1183 #endif
1184 	case ARPHRD_IEEE80211_PRISM:
1185 		handle->linktype = DLT_PRISM_HEADER;
1186 		break;
1187 
1188 #ifndef ARPHRD_IEEE80211_RADIOTAP /* new */
1189 #define ARPHRD_IEEE80211_RADIOTAP 803
1190 #endif
1191 	case ARPHRD_IEEE80211_RADIOTAP:
1192 		handle->linktype = DLT_IEEE802_11_RADIO;
1193 		break;
1194 
1195 	case ARPHRD_PPP:
1196 		/*
1197 		 * Some PPP code in the kernel supplies no link-layer
1198 		 * header whatsoever to PF_PACKET sockets; other PPP
1199 		 * code supplies PPP link-layer headers ("syncppp.c");
1200 		 * some PPP code might supply random link-layer
1201 		 * headers (PPP over ISDN - there's code in Ethereal,
1202 		 * for example, to cope with PPP-over-ISDN captures
1203 		 * with which the Ethereal developers have had to cope,
1204 		 * heuristically trying to determine which of the
1205 		 * oddball link-layer headers particular packets have).
1206 		 *
1207 		 * As such, we just punt, and run all PPP interfaces
1208 		 * in cooked mode, if we can; otherwise, we just treat
1209 		 * it as DLT_RAW, for now - if somebody needs to capture,
1210 		 * on a 2.0[.x] kernel, on PPP devices that supply a
1211 		 * link-layer header, they'll have to add code here to
1212 		 * map to the appropriate DLT_ type (possibly adding a
1213 		 * new DLT_ type, if necessary).
1214 		 */
1215 		if (cooked_ok)
1216 			handle->linktype = DLT_LINUX_SLL;
1217 		else {
1218 			/*
1219 			 * XXX - handle ISDN types here?  We can't fall
1220 			 * back on cooked sockets, so we'd have to
1221 			 * figure out from the device name what type of
1222 			 * link-layer encapsulation it's using, and map
1223 			 * that to an appropriate DLT_ value, meaning
1224 			 * we'd map "isdnN" devices to DLT_RAW (they
1225 			 * supply raw IP packets with no link-layer
1226 			 * header) and "isdY" devices to a new DLT_I4L_IP
1227 			 * type that has only an Ethernet packet type as
1228 			 * a link-layer header.
1229 			 *
1230 			 * But sometimes we seem to get random crap
1231 			 * in the link-layer header when capturing on
1232 			 * ISDN devices....
1233 			 */
1234 			handle->linktype = DLT_RAW;
1235 		}
1236 		break;
1237 
1238 #ifndef ARPHRD_CISCO
1239 #define ARPHRD_CISCO 513 /* previously ARPHRD_HDLC */
1240 #endif
1241 	case ARPHRD_CISCO:
1242 		handle->linktype = DLT_C_HDLC;
1243 		break;
1244 
1245 	/* Not sure if this is correct for all tunnels, but it
1246 	 * works for CIPE */
1247 	case ARPHRD_TUNNEL:
1248 #ifndef ARPHRD_SIT
1249 #define ARPHRD_SIT 776	/* From Linux 2.2.13 */
1250 #endif
1251 	case ARPHRD_SIT:
1252 	case ARPHRD_CSLIP:
1253 	case ARPHRD_SLIP6:
1254 	case ARPHRD_CSLIP6:
1255 	case ARPHRD_ADAPT:
1256 	case ARPHRD_SLIP:
1257 #ifndef ARPHRD_RAWHDLC
1258 #define ARPHRD_RAWHDLC 518
1259 #endif
1260 	case ARPHRD_RAWHDLC:
1261 #ifndef ARPHRD_DLCI
1262 #define ARPHRD_DLCI 15
1263 #endif
1264 	case ARPHRD_DLCI:
1265 		/*
1266 		 * XXX - should some of those be mapped to DLT_LINUX_SLL
1267 		 * instead?  Should we just map all of them to DLT_LINUX_SLL?
1268 		 */
1269 		handle->linktype = DLT_RAW;
1270 		break;
1271 
1272 #ifndef ARPHRD_FRAD
1273 #define ARPHRD_FRAD 770
1274 #endif
1275 	case ARPHRD_FRAD:
1276 		handle->linktype = DLT_FRELAY;
1277 		break;
1278 
1279 	case ARPHRD_LOCALTLK:
1280 		handle->linktype = DLT_LTALK;
1281 		break;
1282 
1283 #ifndef ARPHRD_FCPP
1284 #define ARPHRD_FCPP	784
1285 #endif
1286 	case ARPHRD_FCPP:
1287 #ifndef ARPHRD_FCAL
1288 #define ARPHRD_FCAL	785
1289 #endif
1290 	case ARPHRD_FCAL:
1291 #ifndef ARPHRD_FCPL
1292 #define ARPHRD_FCPL	786
1293 #endif
1294 	case ARPHRD_FCPL:
1295 #ifndef ARPHRD_FCFABRIC
1296 #define ARPHRD_FCFABRIC	787
1297 #endif
1298 	case ARPHRD_FCFABRIC:
1299 		/*
1300 		 * We assume that those all mean RFC 2625 IP-over-
1301 		 * Fibre Channel, with the RFC 2625 header at
1302 		 * the beginning of the packet.
1303 		 */
1304 		handle->linktype = DLT_IP_OVER_FC;
1305 		break;
1306 
1307 #ifndef ARPHRD_IRDA
1308 #define ARPHRD_IRDA	783
1309 #endif
1310 	case ARPHRD_IRDA:
1311 		/* Don't expect IP packet out of this interfaces... */
1312 		handle->linktype = DLT_LINUX_IRDA;
1313 		/* We need to save packet direction for IrDA decoding,
1314 		 * so let's use "Linux-cooked" mode. Jean II */
1315 		//handle->md.cooked = 1;
1316 		break;
1317 
1318 	default:
1319 		handle->linktype = -1;
1320 		break;
1321 	}
1322 }
1323 
1324 /* ===== Functions to interface to the newer kernels ================== */
1325 
1326 /*
1327  *  Try to open a packet socket using the new kernel interface.
1328  *  Returns 0 on failure.
1329  *  FIXME: 0 uses to mean success (Sebastian)
1330  */
1331 static int
1332 live_open_new(pcap_t *handle, const char *device, int promisc,
1333 	      int to_ms, char *ebuf)
1334 {
1335 #ifdef HAVE_PF_PACKET_SOCKETS
1336 	int			sock_fd = -1, arptype;
1337 	int			err;
1338 	int			fatal_err = 0;
1339 	struct packet_mreq	mr;
1340 
1341 	/* One shot loop used for error handling - bail out with break */
1342 
1343 	do {
1344 		/*
1345 		 * Open a socket with protocol family packet. If a device is
1346 		 * given we try to open it in raw mode otherwise we use
1347 		 * the cooked interface.
1348 		 */
1349 		sock_fd = device ?
1350 			socket(PF_PACKET, SOCK_RAW, htons(ETH_P_ALL))
1351 		      : socket(PF_PACKET, SOCK_DGRAM, htons(ETH_P_ALL));
1352 
1353 		if (sock_fd == -1) {
1354 			snprintf(ebuf, PCAP_ERRBUF_SIZE, "socket: %s",
1355 				 pcap_strerror(errno) );
1356 			break;
1357 		}
1358 
1359 		/* It seems the kernel supports the new interface. */
1360 		handle->md.sock_packet = 0;
1361 
1362 		/*
1363 		 * Get the interface index of the loopback device.
1364 		 * If the attempt fails, don't fail, just set the
1365 		 * "md.lo_ifindex" to -1.
1366 		 *
1367 		 * XXX - can there be more than one device that loops
1368 		 * packets back, i.e. devices other than "lo"?  If so,
1369 		 * we'd need to find them all, and have an array of
1370 		 * indices for them, and check all of them in
1371 		 * "pcap_read_packet()".
1372 		 */
1373 		handle->md.lo_ifindex = iface_get_id(sock_fd, "lo", ebuf);
1374 
1375 		/*
1376 		 * Default value for offset to align link-layer payload
1377 		 * on a 4-byte boundary.
1378 		 */
1379 		handle->offset	 = 0;
1380 
1381 		/*
1382 		 * What kind of frames do we have to deal with? Fall back
1383 		 * to cooked mode if we have an unknown interface type.
1384 		 */
1385 
1386 		if (device) {
1387 			/* Assume for now we don't need cooked mode. */
1388 			handle->md.cooked = 0;
1389 
1390 			arptype	= iface_get_arptype(sock_fd, device, ebuf);
1391 			if (arptype == -1) {
1392 				fatal_err = 1;
1393 				break;
1394 			}
1395 			map_arphrd_to_dlt(handle, arptype, 1);
1396 			if (handle->linktype == -1 ||
1397 			    handle->linktype == DLT_LINUX_SLL ||
1398 			    handle->linktype == DLT_LINUX_IRDA ||
1399 			    (handle->linktype == DLT_EN10MB &&
1400 			     (strncmp("isdn", device, 4) == 0 ||
1401 			      strncmp("isdY", device, 4) == 0))) {
1402 				/*
1403 				 * Unknown interface type (-1), or a
1404 				 * device we explicitly chose to run
1405 				 * in cooked mode (e.g., PPP devices),
1406 				 * or an ISDN device (whose link-layer
1407 				 * type we can only determine by using
1408 				 * APIs that may be different on different
1409 				 * kernels) - reopen in cooked mode.
1410 				 */
1411 				if (close(sock_fd) == -1) {
1412 					snprintf(ebuf, PCAP_ERRBUF_SIZE,
1413 						 "close: %s", pcap_strerror(errno));
1414 					break;
1415 				}
1416 				sock_fd = socket(PF_PACKET, SOCK_DGRAM,
1417 						 htons(ETH_P_ALL));
1418 				if (sock_fd == -1) {
1419 					snprintf(ebuf, PCAP_ERRBUF_SIZE,
1420 						 "socket: %s", pcap_strerror(errno));
1421 					break;
1422 				}
1423 				handle->md.cooked = 1;
1424 
1425 				/*
1426 				 * Get rid of any link-layer type list
1427 				 * we allocated - this only supports cooked
1428 				 * capture.
1429 				 */
1430 				if (handle->dlt_list != NULL) {
1431 					free(handle->dlt_list);
1432 					handle->dlt_list = NULL;
1433 					handle->dlt_count = 0;
1434 				}
1435 
1436 				if (handle->linktype == -1) {
1437 					/*
1438 					 * Warn that we're falling back on
1439 					 * cooked mode; we may want to
1440 					 * update "map_arphrd_to_dlt()"
1441 					 * to handle the new type.
1442 					 */
1443 					snprintf(ebuf, PCAP_ERRBUF_SIZE,
1444 						"arptype %d not "
1445 						"supported by libpcap - "
1446 						"falling back to cooked "
1447 						"socket",
1448 						arptype);
1449 				}
1450 				/* IrDA capture is not a real "cooked" capture,
1451 				 * it's IrLAP frames, not IP packets. */
1452 				if (handle->linktype != DLT_LINUX_IRDA)
1453 					handle->linktype = DLT_LINUX_SLL;
1454 			}
1455 
1456 			handle->md.ifindex = iface_get_id(sock_fd, device, ebuf);
1457 			if (handle->md.ifindex == -1)
1458 				break;
1459 
1460 			if ((err = iface_bind(sock_fd, handle->md.ifindex,
1461 			    ebuf)) < 0) {
1462 				if (err == -2)
1463 					fatal_err = 1;
1464 				break;
1465 			}
1466 		} else {
1467 			/*
1468 			 * This is cooked mode.
1469 			 */
1470 			handle->md.cooked = 1;
1471 			handle->linktype = DLT_LINUX_SLL;
1472 
1473 			/*
1474 			 * We're not bound to a device.
1475 			 * XXX - true?  Or true only if we're using
1476 			 * the "any" device?
1477 			 * For now, we're using this as an indication
1478 			 * that we can't transmit; stop doing that only
1479 			 * if we figure out how to transmit in cooked
1480 			 * mode.
1481 			 */
1482 			handle->md.ifindex = -1;
1483 		}
1484 
1485 		/*
1486 		 * Select promiscuous mode on if "promisc" is set.
1487 		 *
1488 		 * Do not turn allmulti mode on if we don't select
1489 		 * promiscuous mode - on some devices (e.g., Orinoco
1490 		 * wireless interfaces), allmulti mode isn't supported
1491 		 * and the driver implements it by turning promiscuous
1492 		 * mode on, and that screws up the operation of the
1493 		 * card as a normal networking interface, and on no
1494 		 * other platform I know of does starting a non-
1495 		 * promiscuous capture affect which multicast packets
1496 		 * are received by the interface.
1497 		 */
1498 
1499 		/*
1500 		 * Hmm, how can we set promiscuous mode on all interfaces?
1501 		 * I am not sure if that is possible at all.
1502 		 */
1503 
1504 		if (device && promisc) {
1505 			memset(&mr, 0, sizeof(mr));
1506 			mr.mr_ifindex = handle->md.ifindex;
1507 			mr.mr_type    = PACKET_MR_PROMISC;
1508 			if (setsockopt(sock_fd, SOL_PACKET,
1509 				PACKET_ADD_MEMBERSHIP, &mr, sizeof(mr)) == -1)
1510 			{
1511 				snprintf(ebuf, PCAP_ERRBUF_SIZE,
1512 					"setsockopt: %s", pcap_strerror(errno));
1513 				break;
1514 			}
1515 		}
1516 
1517 		/* Save the socket FD in the pcap structure */
1518 
1519 		handle->fd 	 = sock_fd;
1520 
1521 		return 1;
1522 
1523 	} while(0);
1524 
1525 	if (sock_fd != -1)
1526 		close(sock_fd);
1527 
1528 	if (fatal_err) {
1529 		/*
1530 		 * Get rid of any link-layer type list we allocated.
1531 		 */
1532 		if (handle->dlt_list != NULL)
1533 			free(handle->dlt_list);
1534 		return -2;
1535 	} else
1536 		return 0;
1537 #else
1538 	strncpy(ebuf,
1539 		"New packet capturing interface not supported by build "
1540 		"environment", PCAP_ERRBUF_SIZE);
1541 	return 0;
1542 #endif
1543 }
1544 
1545 #ifdef HAVE_PF_PACKET_SOCKETS
1546 /*
1547  *  Return the index of the given device name. Fill ebuf and return
1548  *  -1 on failure.
1549  */
1550 static int
1551 iface_get_id(int fd, const char *device, char *ebuf)
1552 {
1553 	struct ifreq	ifr;
1554 
1555 	memset(&ifr, 0, sizeof(ifr));
1556 	strncpy(ifr.ifr_name, device, sizeof(ifr.ifr_name));
1557 
1558 	if (ioctl(fd, SIOCGIFINDEX, &ifr) == -1) {
1559 		snprintf(ebuf, PCAP_ERRBUF_SIZE,
1560 			 "ioctl: %s", pcap_strerror(errno));
1561 		return -1;
1562 	}
1563 
1564 	return ifr.ifr_ifindex;
1565 }
1566 
1567 /*
1568  *  Bind the socket associated with FD to the given device.
1569  */
1570 static int
1571 iface_bind(int fd, int ifindex, char *ebuf)
1572 {
1573 	struct sockaddr_ll	sll;
1574 	int			err;
1575 	socklen_t		errlen = sizeof(err);
1576 
1577 	memset(&sll, 0, sizeof(sll));
1578 	sll.sll_family		= AF_PACKET;
1579 	sll.sll_ifindex		= ifindex;
1580 	sll.sll_protocol	= htons(ETH_P_ALL);
1581 
1582 	if (bind(fd, (struct sockaddr *) &sll, sizeof(sll)) == -1) {
1583 		snprintf(ebuf, PCAP_ERRBUF_SIZE,
1584 			 "bind: %s", pcap_strerror(errno));
1585 		return -1;
1586 	}
1587 
1588 	/* Any pending errors, e.g., network is down? */
1589 
1590 	if (getsockopt(fd, SOL_SOCKET, SO_ERROR, &err, &errlen) == -1) {
1591 		snprintf(ebuf, PCAP_ERRBUF_SIZE,
1592 			"getsockopt: %s", pcap_strerror(errno));
1593 		return -2;
1594 	}
1595 
1596 	if (err > 0) {
1597 		snprintf(ebuf, PCAP_ERRBUF_SIZE,
1598 			"bind: %s", pcap_strerror(err));
1599 		return -2;
1600 	}
1601 
1602 	return 0;
1603 }
1604 
1605 #endif
1606 
1607 
1608 /* ===== Functions to interface to the older kernels ================== */
1609 
1610 /*
1611  * With older kernels promiscuous mode is kind of interesting because we
1612  * have to reset the interface before exiting. The problem can't really
1613  * be solved without some daemon taking care of managing usage counts.
1614  * If we put the interface into promiscuous mode, we set a flag indicating
1615  * that we must take it out of that mode when the interface is closed,
1616  * and, when closing the interface, if that flag is set we take it out
1617  * of promiscuous mode.
1618  */
1619 
1620 /*
1621  * List of pcaps for which we turned promiscuous mode on by hand.
1622  * If there are any such pcaps, we arrange to call "pcap_close_all()"
1623  * when we exit, and have it close all of them to turn promiscuous mode
1624  * off.
1625  */
1626 static struct pcap *pcaps_to_close;
1627 
1628 /*
1629  * TRUE if we've already called "atexit()" to cause "pcap_close_all()" to
1630  * be called on exit.
1631  */
1632 static int did_atexit;
1633 
1634 static void	pcap_close_all(void)
1635 {
1636 	struct pcap *handle;
1637 
1638 	while ((handle = pcaps_to_close) != NULL)
1639 		pcap_close(handle);
1640 }
1641 
1642 static void	pcap_close_linux( pcap_t *handle )
1643 {
1644 	struct pcap	*p, *prevp;
1645 	struct ifreq	ifr;
1646 
1647 	if (handle->md.clear_promisc) {
1648 		/*
1649 		 * We put the interface into promiscuous mode; take
1650 		 * it out of promiscuous mode.
1651 		 *
1652 		 * XXX - if somebody else wants it in promiscuous mode,
1653 		 * this code cannot know that, so it'll take it out
1654 		 * of promiscuous mode.  That's not fixable in 2.0[.x]
1655 		 * kernels.
1656 		 */
1657 		memset(&ifr, 0, sizeof(ifr));
1658 		strncpy(ifr.ifr_name, handle->md.device, sizeof(ifr.ifr_name));
1659 		if (ioctl(handle->fd, SIOCGIFFLAGS, &ifr) == -1) {
1660 			fprintf(stderr,
1661 			    "Can't restore interface flags (SIOCGIFFLAGS failed: %s).\n"
1662 			    "Please adjust manually.\n"
1663 			    "Hint: This can't happen with Linux >= 2.2.0.\n",
1664 			    strerror(errno));
1665 		} else {
1666 			if (ifr.ifr_flags & IFF_PROMISC) {
1667 				/*
1668 				 * Promiscuous mode is currently on; turn it
1669 				 * off.
1670 				 */
1671 				ifr.ifr_flags &= ~IFF_PROMISC;
1672 				if (ioctl(handle->fd, SIOCSIFFLAGS, &ifr) == -1) {
1673 					fprintf(stderr,
1674 					    "Can't restore interface flags (SIOCSIFFLAGS failed: %s).\n"
1675 					    "Please adjust manually.\n"
1676 					    "Hint: This can't happen with Linux >= 2.2.0.\n",
1677 					    strerror(errno));
1678 				}
1679 			}
1680 		}
1681 
1682 		/*
1683 		 * Take this pcap out of the list of pcaps for which we
1684 		 * have to take the interface out of promiscuous mode.
1685 		 */
1686 		for (p = pcaps_to_close, prevp = NULL; p != NULL;
1687 		    prevp = p, p = p->md.next) {
1688 			if (p == handle) {
1689 				/*
1690 				 * Found it.  Remove it from the list.
1691 				 */
1692 				if (prevp == NULL) {
1693 					/*
1694 					 * It was at the head of the list.
1695 					 */
1696 					pcaps_to_close = p->md.next;
1697 				} else {
1698 					/*
1699 					 * It was in the middle of the list.
1700 					 */
1701 					prevp->md.next = p->md.next;
1702 				}
1703 				break;
1704 			}
1705 		}
1706 	}
1707 
1708 	if (handle->md.device != NULL)
1709 		free(handle->md.device);
1710 	handle->md.device = NULL;
1711 	pcap_close_common(handle);
1712 }
1713 
1714 /*
1715  *  Try to open a packet socket using the old kernel interface.
1716  *  Returns 0 on failure.
1717  *  FIXME: 0 uses to mean success (Sebastian)
1718  */
1719 static int
1720 live_open_old(pcap_t *handle, const char *device, int promisc,
1721 	      int to_ms, char *ebuf)
1722 {
1723 	int		arptype;
1724 	struct ifreq	ifr;
1725 
1726 	do {
1727 		/* Open the socket */
1728 
1729 		handle->fd = socket(PF_INET, SOCK_PACKET, htons(ETH_P_ALL));
1730 		if (handle->fd == -1) {
1731 			snprintf(ebuf, PCAP_ERRBUF_SIZE,
1732 				 "socket: %s", pcap_strerror(errno));
1733 			break;
1734 		}
1735 
1736 		/* It worked - we are using the old interface */
1737 		handle->md.sock_packet = 1;
1738 
1739 		/* ...which means we get the link-layer header. */
1740 		handle->md.cooked = 0;
1741 
1742 		/* Bind to the given device */
1743 
1744 		if (!device) {
1745 		        strncpy(ebuf, "pcap_open_live: The \"any\" device isn't supported on 2.0[.x]-kernel systems",
1746 				PCAP_ERRBUF_SIZE);
1747 			break;
1748 		}
1749 		if (iface_bind_old(handle->fd, device, ebuf) == -1)
1750 			break;
1751 
1752 		/*
1753 		 * Try to get the link-layer type.
1754 		 */
1755 		arptype = iface_get_arptype(handle->fd, device, ebuf);
1756 		if (arptype == -1)
1757 			break;
1758 
1759 		/*
1760 		 * Try to find the DLT_ type corresponding to that
1761 		 * link-layer type.
1762 		 */
1763 		map_arphrd_to_dlt(handle, arptype, 0);
1764 		if (handle->linktype == -1) {
1765 			snprintf(ebuf, PCAP_ERRBUF_SIZE,
1766 				 "unknown arptype %d", arptype);
1767 			break;
1768 		}
1769 
1770 		/* Go to promisc mode if requested */
1771 
1772 		if (promisc) {
1773 			memset(&ifr, 0, sizeof(ifr));
1774 			strncpy(ifr.ifr_name, device, sizeof(ifr.ifr_name));
1775 			if (ioctl(handle->fd, SIOCGIFFLAGS, &ifr) == -1) {
1776 				snprintf(ebuf, PCAP_ERRBUF_SIZE,
1777 					 "ioctl: %s", pcap_strerror(errno));
1778 				break;
1779 			}
1780 			if ((ifr.ifr_flags & IFF_PROMISC) == 0) {
1781 				/*
1782 				 * Promiscuous mode isn't currently on,
1783 				 * so turn it on, and remember that
1784 				 * we should turn it off when the
1785 				 * pcap_t is closed.
1786 				 */
1787 
1788 				/*
1789 				 * If we haven't already done so, arrange
1790 				 * to have "pcap_close_all()" called when
1791 				 * we exit.
1792 				 */
1793 				if (!did_atexit) {
1794 					if (atexit(pcap_close_all) == -1) {
1795 						/*
1796 						 * "atexit()" failed; don't
1797 						 * put the interface in
1798 						 * promiscuous mode, just
1799 						 * give up.
1800 						 */
1801 						strncpy(ebuf, "atexit failed",
1802 							PCAP_ERRBUF_SIZE);
1803 						break;
1804 					}
1805 					did_atexit = 1;
1806 				}
1807 
1808 				ifr.ifr_flags |= IFF_PROMISC;
1809 				if (ioctl(handle->fd, SIOCSIFFLAGS, &ifr) == -1) {
1810 				        snprintf(ebuf, PCAP_ERRBUF_SIZE,
1811 						 "ioctl: %s",
1812 						 pcap_strerror(errno));
1813 					break;
1814 				}
1815 				handle->md.clear_promisc = 1;
1816 
1817 				/*
1818 				 * Add this to the list of pcaps
1819 				 * to close when we exit.
1820 				 */
1821 				handle->md.next = pcaps_to_close;
1822 				pcaps_to_close = handle;
1823 			}
1824 		}
1825 
1826 		/*
1827 		 * Default value for offset to align link-layer payload
1828 		 * on a 4-byte boundary.
1829 		 */
1830 		handle->offset	 = 0;
1831 
1832 		return 1;
1833 
1834 	} while (0);
1835 
1836 	pcap_close_linux(handle);
1837 	return 0;
1838 }
1839 
1840 /*
1841  *  Bind the socket associated with FD to the given device using the
1842  *  interface of the old kernels.
1843  */
1844 static int
1845 iface_bind_old(int fd, const char *device, char *ebuf)
1846 {
1847 	struct sockaddr	saddr;
1848 	int		err;
1849 	socklen_t	errlen = sizeof(err);
1850 
1851 	memset(&saddr, 0, sizeof(saddr));
1852 	strncpy(saddr.sa_data, device, sizeof(saddr.sa_data));
1853 	if (bind(fd, &saddr, sizeof(saddr)) == -1) {
1854 		snprintf(ebuf, PCAP_ERRBUF_SIZE,
1855 			 "bind: %s", pcap_strerror(errno));
1856 		return -1;
1857 	}
1858 
1859 	/* Any pending errors, e.g., network is down? */
1860 
1861 	if (getsockopt(fd, SOL_SOCKET, SO_ERROR, &err, &errlen) == -1) {
1862 		snprintf(ebuf, PCAP_ERRBUF_SIZE,
1863 			"getsockopt: %s", pcap_strerror(errno));
1864 		return -1;
1865 	}
1866 
1867 	if (err > 0) {
1868 		snprintf(ebuf, PCAP_ERRBUF_SIZE,
1869 			"bind: %s", pcap_strerror(err));
1870 		return -1;
1871 	}
1872 
1873 	return 0;
1874 }
1875 
1876 
1877 /* ===== System calls available on all supported kernels ============== */
1878 
1879 /*
1880  *  Query the kernel for the MTU of the given interface.
1881  */
1882 static int
1883 iface_get_mtu(int fd, const char *device, char *ebuf)
1884 {
1885 	struct ifreq	ifr;
1886 
1887 	if (!device)
1888 		return BIGGER_THAN_ALL_MTUS;
1889 
1890 	memset(&ifr, 0, sizeof(ifr));
1891 	strncpy(ifr.ifr_name, device, sizeof(ifr.ifr_name));
1892 
1893 	if (ioctl(fd, SIOCGIFMTU, &ifr) == -1) {
1894 		snprintf(ebuf, PCAP_ERRBUF_SIZE,
1895 			 "ioctl: %s", pcap_strerror(errno));
1896 		return -1;
1897 	}
1898 
1899 	return ifr.ifr_mtu;
1900 }
1901 
1902 /*
1903  *  Get the hardware type of the given interface as ARPHRD_xxx constant.
1904  */
1905 static int
1906 iface_get_arptype(int fd, const char *device, char *ebuf)
1907 {
1908 	struct ifreq	ifr;
1909 
1910 	memset(&ifr, 0, sizeof(ifr));
1911 	strncpy(ifr.ifr_name, device, sizeof(ifr.ifr_name));
1912 
1913 	if (ioctl(fd, SIOCGIFHWADDR, &ifr) == -1) {
1914 		snprintf(ebuf, PCAP_ERRBUF_SIZE,
1915 			 "ioctl: %s", pcap_strerror(errno));
1916 		return -1;
1917 	}
1918 
1919 	return ifr.ifr_hwaddr.sa_family;
1920 }
1921 
1922 #ifdef SO_ATTACH_FILTER
1923 static int
1924 fix_program(pcap_t *handle, struct sock_fprog *fcode)
1925 {
1926 	size_t prog_size;
1927 	register int i;
1928 	register struct bpf_insn *p;
1929 	struct bpf_insn *f;
1930 	int len;
1931 
1932 	/*
1933 	 * Make a copy of the filter, and modify that copy if
1934 	 * necessary.
1935 	 */
1936 	prog_size = sizeof(*handle->fcode.bf_insns) * handle->fcode.bf_len;
1937 	len = handle->fcode.bf_len;
1938 	f = (struct bpf_insn *)malloc(prog_size);
1939 	if (f == NULL) {
1940 		snprintf(handle->errbuf, sizeof(handle->errbuf),
1941 			 "malloc: %s", pcap_strerror(errno));
1942 		return -1;
1943 	}
1944 	memcpy(f, handle->fcode.bf_insns, prog_size);
1945 	fcode->len = len;
1946 	fcode->filter = (struct sock_filter *) f;
1947 
1948 	for (i = 0; i < len; ++i) {
1949 		p = &f[i];
1950 		/*
1951 		 * What type of instruction is this?
1952 		 */
1953 		switch (BPF_CLASS(p->code)) {
1954 
1955 		case BPF_RET:
1956 			/*
1957 			 * It's a return instruction; is the snapshot
1958 			 * length a constant, rather than the contents
1959 			 * of the accumulator?
1960 			 */
1961 			if (BPF_MODE(p->code) == BPF_K) {
1962 				/*
1963 				 * Yes - if the value to be returned,
1964 				 * i.e. the snapshot length, is anything
1965 				 * other than 0, make it 65535, so that
1966 				 * the packet is truncated by "recvfrom()",
1967 				 * not by the filter.
1968 				 *
1969 				 * XXX - there's nothing we can easily do
1970 				 * if it's getting the value from the
1971 				 * accumulator; we'd have to insert
1972 				 * code to force non-zero values to be
1973 				 * 65535.
1974 				 */
1975 				if (p->k != 0)
1976 					p->k = 65535;
1977 			}
1978 			break;
1979 
1980 		case BPF_LD:
1981 		case BPF_LDX:
1982 			/*
1983 			 * It's a load instruction; is it loading
1984 			 * from the packet?
1985 			 */
1986 			switch (BPF_MODE(p->code)) {
1987 
1988 			case BPF_ABS:
1989 			case BPF_IND:
1990 			case BPF_MSH:
1991 				/*
1992 				 * Yes; are we in cooked mode?
1993 				 */
1994 				if (handle->md.cooked) {
1995 					/*
1996 					 * Yes, so we need to fix this
1997 					 * instruction.
1998 					 */
1999 					if (fix_offset(p) < 0) {
2000 						/*
2001 						 * We failed to do so.
2002 						 * Return 0, so our caller
2003 						 * knows to punt to userland.
2004 						 */
2005 						return 0;
2006 					}
2007 				}
2008 				break;
2009 			}
2010 			break;
2011 		}
2012 	}
2013 	return 1;	/* we succeeded */
2014 }
2015 
2016 static int
2017 fix_offset(struct bpf_insn *p)
2018 {
2019 	/*
2020 	 * What's the offset?
2021 	 */
2022 	if (p->k >= SLL_HDR_LEN) {
2023 		/*
2024 		 * It's within the link-layer payload; that starts at an
2025 		 * offset of 0, as far as the kernel packet filter is
2026 		 * concerned, so subtract the length of the link-layer
2027 		 * header.
2028 		 */
2029 		p->k -= SLL_HDR_LEN;
2030 	} else if (p->k == 14) {
2031 		/*
2032 		 * It's the protocol field; map it to the special magic
2033 		 * kernel offset for that field.
2034 		 */
2035 		p->k = SKF_AD_OFF + SKF_AD_PROTOCOL;
2036 	} else {
2037 		/*
2038 		 * It's within the header, but it's not one of those
2039 		 * fields; we can't do that in the kernel, so punt
2040 		 * to userland.
2041 		 */
2042 		return -1;
2043 	}
2044 	return 0;
2045 }
2046 
2047 static int
2048 set_kernel_filter(pcap_t *handle, struct sock_fprog *fcode)
2049 {
2050 	int total_filter_on = 0;
2051 	int save_mode;
2052 	int ret;
2053 	int save_errno;
2054 
2055 	/*
2056 	 * The socket filter code doesn't discard all packets queued
2057 	 * up on the socket when the filter is changed; this means
2058 	 * that packets that don't match the new filter may show up
2059 	 * after the new filter is put onto the socket, if those
2060 	 * packets haven't yet been read.
2061 	 *
2062 	 * This means, for example, that if you do a tcpdump capture
2063 	 * with a filter, the first few packets in the capture might
2064 	 * be packets that wouldn't have passed the filter.
2065 	 *
2066 	 * We therefore discard all packets queued up on the socket
2067 	 * when setting a kernel filter.  (This isn't an issue for
2068 	 * userland filters, as the userland filtering is done after
2069 	 * packets are queued up.)
2070 	 *
2071 	 * To flush those packets, we put the socket in read-only mode,
2072 	 * and read packets from the socket until there are no more to
2073 	 * read.
2074 	 *
2075 	 * In order to keep that from being an infinite loop - i.e.,
2076 	 * to keep more packets from arriving while we're draining
2077 	 * the queue - we put the "total filter", which is a filter
2078 	 * that rejects all packets, onto the socket before draining
2079 	 * the queue.
2080 	 *
2081 	 * This code deliberately ignores any errors, so that you may
2082 	 * get bogus packets if an error occurs, rather than having
2083 	 * the filtering done in userland even if it could have been
2084 	 * done in the kernel.
2085 	 */
2086 	if (setsockopt(handle->fd, SOL_SOCKET, SO_ATTACH_FILTER,
2087 		       &total_fcode, sizeof(total_fcode)) == 0) {
2088 		char drain[1];
2089 
2090 		/*
2091 		 * Note that we've put the total filter onto the socket.
2092 		 */
2093 		total_filter_on = 1;
2094 
2095 		/*
2096 		 * Save the socket's current mode, and put it in
2097 		 * non-blocking mode; we drain it by reading packets
2098 		 * until we get an error (which is normally a
2099 		 * "nothing more to be read" error).
2100 		 */
2101 		save_mode = fcntl(handle->fd, F_GETFL, 0);
2102 		if (save_mode != -1 &&
2103 		    fcntl(handle->fd, F_SETFL, save_mode | O_NONBLOCK) >= 0) {
2104 			while (recv(handle->fd, &drain, sizeof drain,
2105 			       MSG_TRUNC) >= 0)
2106 				;
2107 			save_errno = errno;
2108 			fcntl(handle->fd, F_SETFL, save_mode);
2109 			if (save_errno != EAGAIN) {
2110 				/* Fatal error */
2111 				reset_kernel_filter(handle);
2112 				snprintf(handle->errbuf, sizeof(handle->errbuf),
2113 				 "recv: %s", pcap_strerror(save_errno));
2114 				return -2;
2115 			}
2116 		}
2117 	}
2118 
2119 	/*
2120 	 * Now attach the new filter.
2121 	 */
2122 	ret = setsockopt(handle->fd, SOL_SOCKET, SO_ATTACH_FILTER,
2123 			 fcode, sizeof(*fcode));
2124 	if (ret == -1 && total_filter_on) {
2125 		/*
2126 		 * Well, we couldn't set that filter on the socket,
2127 		 * but we could set the total filter on the socket.
2128 		 *
2129 		 * This could, for example, mean that the filter was
2130 		 * too big to put into the kernel, so we'll have to
2131 		 * filter in userland; in any case, we'll be doing
2132 		 * filtering in userland, so we need to remove the
2133 		 * total filter so we see packets.
2134 		 */
2135 		save_errno = errno;
2136 
2137 		/*
2138 		 * XXX - if this fails, we're really screwed;
2139 		 * we have the total filter on the socket,
2140 		 * and it won't come off.  What do we do then?
2141 		 */
2142 		reset_kernel_filter(handle);
2143 
2144 		errno = save_errno;
2145 	}
2146 	return ret;
2147 }
2148 
2149 static int
2150 reset_kernel_filter(pcap_t *handle)
2151 {
2152 	/* setsockopt() barfs unless it get a dummy parameter */
2153 	int dummy;
2154 
2155 	return setsockopt(handle->fd, SOL_SOCKET, SO_DETACH_FILTER,
2156 				   &dummy, sizeof(dummy));
2157 }
2158 #endif
2159