xref: /freebsd/share/man/man4/bpf.4 (revision 4f1f4356f3012928b463f9ef1710fb908e48b1e2)
1.\" Copyright (c) 2007 Seccuris Inc.
2.\" All rights reserved.
3.\"
4.\" This software was developed by Robert N. M. Watson under contract to
5.\" Seccuris Inc.
6.\"
7.\" Redistribution and use in source and binary forms, with or without
8.\" modification, are permitted provided that the following conditions
9.\" are met:
10.\" 1. Redistributions of source code must retain the above copyright
11.\"    notice, this list of conditions and the following disclaimer.
12.\" 2. Redistributions in binary form must reproduce the above copyright
13.\"    notice, this list of conditions and the following disclaimer in the
14.\"    documentation and/or other materials provided with the distribution.
15.\"
16.\" THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
17.\" ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
18.\" IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
19.\" ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
20.\" FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
21.\" DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
22.\" OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
23.\" HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
24.\" LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
25.\" OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
26.\" SUCH DAMAGE.
27.\"
28.\" Copyright (c) 1990 The Regents of the University of California.
29.\" All rights reserved.
30.\"
31.\" Redistribution and use in source and binary forms, with or without
32.\" modification, are permitted provided that: (1) source code distributions
33.\" retain the above copyright notice and this paragraph in its entirety, (2)
34.\" distributions including binary code include the above copyright notice and
35.\" this paragraph in its entirety in the documentation or other materials
36.\" provided with the distribution, and (3) all advertising materials mentioning
37.\" features or use of this software display the following acknowledgement:
38.\" ``This product includes software developed by the University of California,
39.\" Lawrence Berkeley Laboratory and its contributors.'' Neither the name of
40.\" the University nor the names of its contributors may be used to endorse
41.\" or promote products derived from this software without specific prior
42.\" written permission.
43.\" THIS SOFTWARE IS PROVIDED ``AS IS'' AND WITHOUT ANY EXPRESS OR IMPLIED
44.\" WARRANTIES, INCLUDING, WITHOUT LIMITATION, THE IMPLIED WARRANTIES OF
45.\" MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE.
46.\"
47.\" This document is derived in part from the enet man page (enet.4)
48.\" distributed with 4.3BSD Unix.
49.\"
50.\" $FreeBSD$
51.\"
52.Dd June 15, 2010
53.Dt BPF 4
54.Os
55.Sh NAME
56.Nm bpf
57.Nd Berkeley Packet Filter
58.Sh SYNOPSIS
59.Cd device bpf
60.Sh DESCRIPTION
61The Berkeley Packet Filter
62provides a raw interface to data link layers in a protocol
63independent fashion.
64All packets on the network, even those destined for other hosts,
65are accessible through this mechanism.
66.Pp
67The packet filter appears as a character special device,
68.Pa /dev/bpf .
69After opening the device, the file descriptor must be bound to a
70specific network interface with the
71.Dv BIOCSETIF
72ioctl.
73A given interface can be shared by multiple listeners, and the filter
74underlying each descriptor will see an identical packet stream.
75.Pp
76A separate device file is required for each minor device.
77If a file is in use, the open will fail and
78.Va errno
79will be set to
80.Er EBUSY .
81.Pp
82Associated with each open instance of a
83.Nm
84file is a user-settable packet filter.
85Whenever a packet is received by an interface,
86all file descriptors listening on that interface apply their filter.
87Each descriptor that accepts the packet receives its own copy.
88.Pp
89The packet filter will support any link level protocol that has fixed length
90headers.
91Currently, only Ethernet,
92.Tn SLIP ,
93and
94.Tn PPP
95drivers have been modified to interact with
96.Nm .
97.Pp
98Since packet data is in network byte order, applications should use the
99.Xr byteorder 3
100macros to extract multi-byte values.
101.Pp
102A packet can be sent out on the network by writing to a
103.Nm
104file descriptor.
105The writes are unbuffered, meaning only one packet can be processed per write.
106Currently, only writes to Ethernets and
107.Tn SLIP
108links are supported.
109.Sh BUFFER MODES
110.Nm
111devices deliver packet data to the application via memory buffers provided by
112the application.
113The buffer mode is set using the
114.Dv BIOCSETBUFMODE
115ioctl, and read using the
116.Dv BIOCGETBUFMODE
117ioctl.
118.Ss Buffered read mode
119By default,
120.Nm
121devices operate in the
122.Dv BPF_BUFMODE_BUFFER
123mode, in which packet data is copied explicitly from kernel to user memory
124using the
125.Xr read 2
126system call.
127The user process will declare a fixed buffer size that will be used both for
128sizing internal buffers and for all
129.Xr read 2
130operations on the file.
131This size is queried using the
132.Dv BIOCGBLEN
133ioctl, and is set using the
134.Dv BIOCSBLEN
135ioctl.
136Note that an individual packet larger than the buffer size is necessarily
137truncated.
138.Ss Zero-copy buffer mode
139.Nm
140devices may also operate in the
141.Dv BPF_BUFMODE_ZEROCOPY
142mode, in which packet data is written directly into two user memory buffers
143by the kernel, avoiding both system call and copying overhead.
144Buffers are of fixed (and equal) size, page-aligned, and an even multiple of
145the page size.
146The maximum zero-copy buffer size is returned by the
147.Dv BIOCGETZMAX
148ioctl.
149Note that an individual packet larger than the buffer size is necessarily
150truncated.
151.Pp
152The user process registers two memory buffers using the
153.Dv BIOCSETZBUF
154ioctl, which accepts a
155.Vt struct bpf_zbuf
156pointer as an argument:
157.Bd -literal
158struct bpf_zbuf {
159	void *bz_bufa;
160	void *bz_bufb;
161	size_t bz_buflen;
162};
163.Ed
164.Pp
165.Vt bz_bufa
166is a pointer to the userspace address of the first buffer that will be
167filled, and
168.Vt bz_bufb
169is a pointer to the second buffer.
170.Nm
171will then cycle between the two buffers as they fill and are acknowledged.
172.Pp
173Each buffer begins with a fixed-length header to hold synchronization and
174data length information for the buffer:
175.Bd -literal
176struct bpf_zbuf_header {
177	volatile u_int  bzh_kernel_gen;	/* Kernel generation number. */
178	volatile u_int  bzh_kernel_len;	/* Length of data in the buffer. */
179	volatile u_int  bzh_user_gen;	/* User generation number. */
180	/* ...padding for future use... */
181};
182.Ed
183.Pp
184The header structure of each buffer, including all padding, should be zeroed
185before it is configured using
186.Dv BIOCSETZBUF .
187Remaining space in the buffer will be used by the kernel to store packet
188data, laid out in the same format as with buffered read mode.
189.Pp
190The kernel and the user process follow a simple acknowledgement protocol via
191the buffer header to synchronize access to the buffer: when the header
192generation numbers,
193.Vt bzh_kernel_gen
194and
195.Vt bzh_user_gen ,
196hold the same value, the kernel owns the buffer, and when they differ,
197userspace owns the buffer.
198.Pp
199While the kernel owns the buffer, the contents are unstable and may change
200asynchronously; while the user process owns the buffer, its contents are
201stable and will not be changed until the buffer has been acknowledged.
202.Pp
203Initializing the buffer headers to all 0's before registering the buffer has
204the effect of assigning initial ownership of both buffers to the kernel.
205The kernel signals that a buffer has been assigned to userspace by modifying
206.Vt bzh_kernel_gen ,
207and userspace acknowledges the buffer and returns it to the kernel by setting
208the value of
209.Vt bzh_user_gen
210to the value of
211.Vt bzh_kernel_gen .
212.Pp
213In order to avoid caching and memory re-ordering effects, the user process
214must use atomic operations and memory barriers when checking for and
215acknowledging buffers:
216.Bd -literal
217#include <machine/atomic.h>
218
219/*
220 * Return ownership of a buffer to the kernel for reuse.
221 */
222static void
223buffer_acknowledge(struct bpf_zbuf_header *bzh)
224{
225
226	atomic_store_rel_int(&bzh->bzh_user_gen, bzh->bzh_kernel_gen);
227}
228
229/*
230 * Check whether a buffer has been assigned to userspace by the kernel.
231 * Return true if userspace owns the buffer, and false otherwise.
232 */
233static int
234buffer_check(struct bpf_zbuf_header *bzh)
235{
236
237	return (bzh->bzh_user_gen !=
238	    atomic_load_acq_int(&bzh->bzh_kernel_gen));
239}
240.Ed
241.Pp
242The user process may force the assignment of the next buffer, if any data
243is pending, to userspace using the
244.Dv BIOCROTZBUF
245ioctl.
246This allows the user process to retrieve data in a partially filled buffer
247before the buffer is full, such as following a timeout; the process must
248recheck for buffer ownership using the header generation numbers, as the
249buffer will not be assigned to userspace if no data was present.
250.Pp
251As in the buffered read mode,
252.Xr kqueue 2 ,
253.Xr poll 2 ,
254and
255.Xr select 2
256may be used to sleep awaiting the availability of a completed buffer.
257They will return a readable file descriptor when ownership of the next buffer
258is assigned to user space.
259.Pp
260In the current implementation, the kernel may assign zero, one, or both
261buffers to the user process; however, an earlier implementation maintained
262the invariant that at most one buffer could be assigned to the user process
263at a time.
264In order to both ensure progress and high performance, user processes should
265acknowledge a completely processed buffer as quickly as possible, returning
266it for reuse, and not block waiting on a second buffer while holding another
267buffer.
268.Sh IOCTLS
269The
270.Xr ioctl 2
271command codes below are defined in
272.In net/bpf.h .
273All commands require
274these includes:
275.Bd -literal
276	#include <sys/types.h>
277	#include <sys/time.h>
278	#include <sys/ioctl.h>
279	#include <net/bpf.h>
280.Ed
281.Pp
282Additionally,
283.Dv BIOCGETIF
284and
285.Dv BIOCSETIF
286require
287.In sys/socket.h
288and
289.In net/if.h .
290.Pp
291In addition to
292.Dv FIONREAD
293and
294.Dv SIOCGIFADDR ,
295the following commands may be applied to any open
296.Nm
297file.
298The (third) argument to
299.Xr ioctl 2
300should be a pointer to the type indicated.
301.Bl -tag -width BIOCGETBUFMODE
302.It Dv BIOCGBLEN
303.Pq Li u_int
304Returns the required buffer length for reads on
305.Nm
306files.
307.It Dv BIOCSBLEN
308.Pq Li u_int
309Sets the buffer length for reads on
310.Nm
311files.
312The buffer must be set before the file is attached to an interface
313with
314.Dv BIOCSETIF .
315If the requested buffer size cannot be accommodated, the closest
316allowable size will be set and returned in the argument.
317A read call will result in
318.Er EIO
319if it is passed a buffer that is not this size.
320.It Dv BIOCGDLT
321.Pq Li u_int
322Returns the type of the data link layer underlying the attached interface.
323.Er EINVAL
324is returned if no interface has been specified.
325The device types, prefixed with
326.Dq Li DLT_ ,
327are defined in
328.In net/bpf.h .
329.It Dv BIOCPROMISC
330Forces the interface into promiscuous mode.
331All packets, not just those destined for the local host, are processed.
332Since more than one file can be listening on a given interface,
333a listener that opened its interface non-promiscuously may receive
334packets promiscuously.
335This problem can be remedied with an appropriate filter.
336.It Dv BIOCFLUSH
337Flushes the buffer of incoming packets,
338and resets the statistics that are returned by BIOCGSTATS.
339.It Dv BIOCGETIF
340.Pq Li "struct ifreq"
341Returns the name of the hardware interface that the file is listening on.
342The name is returned in the ifr_name field of
343the
344.Li ifreq
345structure.
346All other fields are undefined.
347.It Dv BIOCSETIF
348.Pq Li "struct ifreq"
349Sets the hardware interface associate with the file.
350This
351command must be performed before any packets can be read.
352The device is indicated by name using the
353.Li ifr_name
354field of the
355.Li ifreq
356structure.
357Additionally, performs the actions of
358.Dv BIOCFLUSH .
359.It Dv BIOCSRTIMEOUT
360.It Dv BIOCGRTIMEOUT
361.Pq Li "struct timeval"
362Set or get the read timeout parameter.
363The argument
364specifies the length of time to wait before timing
365out on a read request.
366This parameter is initialized to zero by
367.Xr open 2 ,
368indicating no timeout.
369.It Dv BIOCGSTATS
370.Pq Li "struct bpf_stat"
371Returns the following structure of packet statistics:
372.Bd -literal
373struct bpf_stat {
374	u_int bs_recv;    /* number of packets received */
375	u_int bs_drop;    /* number of packets dropped */
376};
377.Ed
378.Pp
379The fields are:
380.Bl -hang -offset indent
381.It Li bs_recv
382the number of packets received by the descriptor since opened or reset
383(including any buffered since the last read call);
384and
385.It Li bs_drop
386the number of packets which were accepted by the filter but dropped by the
387kernel because of buffer overflows
388(i.e., the application's reads are not keeping up with the packet traffic).
389.El
390.It Dv BIOCIMMEDIATE
391.Pq Li u_int
392Enable or disable
393.Dq immediate mode ,
394based on the truth value of the argument.
395When immediate mode is enabled, reads return immediately upon packet
396reception.
397Otherwise, a read will block until either the kernel buffer
398becomes full or a timeout occurs.
399This is useful for programs like
400.Xr rarpd 8
401which must respond to messages in real time.
402The default for a new file is off.
403.It Dv BIOCSETF
404.It Dv BIOCSETFNR
405.Pq Li "struct bpf_program"
406Sets the read filter program used by the kernel to discard uninteresting
407packets.
408An array of instructions and its length is passed in using
409the following structure:
410.Bd -literal
411struct bpf_program {
412	int bf_len;
413	struct bpf_insn *bf_insns;
414};
415.Ed
416.Pp
417The filter program is pointed to by the
418.Li bf_insns
419field while its length in units of
420.Sq Li struct bpf_insn
421is given by the
422.Li bf_len
423field.
424See section
425.Sx "FILTER MACHINE"
426for an explanation of the filter language.
427The only difference between
428.Dv BIOCSETF
429and
430.Dv BIOCSETFNR
431is
432.Dv BIOCSETF
433performs the actions of
434.Dv BIOCFLUSH
435while
436.Dv BIOCSETFNR
437does not.
438.It Dv BIOCSETWF
439.Pq Li "struct bpf_program"
440Sets the write filter program used by the kernel to control what type of
441packets can be written to the interface.
442See the
443.Dv BIOCSETF
444command for more
445information on the
446.Nm
447filter program.
448.It Dv BIOCVERSION
449.Pq Li "struct bpf_version"
450Returns the major and minor version numbers of the filter language currently
451recognized by the kernel.
452Before installing a filter, applications must check
453that the current version is compatible with the running kernel.
454Version numbers are compatible if the major numbers match and the application minor
455is less than or equal to the kernel minor.
456The kernel version number is returned in the following structure:
457.Bd -literal
458struct bpf_version {
459        u_short bv_major;
460        u_short bv_minor;
461};
462.Ed
463.Pp
464The current version numbers are given by
465.Dv BPF_MAJOR_VERSION
466and
467.Dv BPF_MINOR_VERSION
468from
469.In net/bpf.h .
470An incompatible filter
471may result in undefined behavior (most likely, an error returned by
472.Fn ioctl
473or haphazard packet matching).
474.It Dv BIOCSHDRCMPLT
475.It Dv BIOCGHDRCMPLT
476.Pq Li u_int
477Set or get the status of the
478.Dq header complete
479flag.
480Set to zero if the link level source address should be filled in automatically
481by the interface output routine.
482Set to one if the link level source
483address will be written, as provided, to the wire.
484This flag is initialized to zero by default.
485.It Dv BIOCSSEESENT
486.It Dv BIOCGSEESENT
487.Pq Li u_int
488These commands are obsolete but left for compatibility.
489Use
490.Dv BIOCSDIRECTION
491and
492.Dv BIOCGDIRECTION
493instead.
494Set or get the flag determining whether locally generated packets on the
495interface should be returned by BPF.
496Set to zero to see only incoming packets on the interface.
497Set to one to see packets originating locally and remotely on the interface.
498This flag is initialized to one by default.
499.It Dv BIOCSDIRECTION
500.It Dv BIOCGDIRECTION
501.Pq Li u_int
502Set or get the setting determining whether incoming, outgoing, or all packets
503on the interface should be returned by BPF.
504Set to
505.Dv BPF_D_IN
506to see only incoming packets on the interface.
507Set to
508.Dv BPF_D_INOUT
509to see packets originating locally and remotely on the interface.
510Set to
511.Dv BPF_D_OUT
512to see only outgoing packets on the interface.
513This setting is initialized to
514.Dv BPF_D_INOUT
515by default.
516.It Dv BIOCSTSTAMP
517.It Dv BIOCGTSTAMP
518.Pq Li u_int
519Set or get format and resolution of the time stamps returned by BPF.
520Set to
521.Dv BPF_T_MICROTIME ,
522.Dv BPF_T_MICROTIME_FAST ,
523.Dv BPF_T_MICROTIME_MONOTONIC ,
524or
525.Dv BPF_T_MICROTIME_MONOTONIC_FAST
526to get time stamps in 64-bit
527.Vt struct timeval
528format.
529Set to
530.Dv BPF_T_NANOTIME ,
531.Dv BPF_T_NANOTIME_FAST ,
532.Dv BPF_T_NANOTIME_MONOTONIC ,
533or
534.Dv BPF_T_NANOTIME_MONOTONIC_FAST
535to get time stamps in 64-bit
536.Vt struct timespec
537format.
538Set to
539.Dv BPF_T_BINTIME ,
540.Dv BPF_T_BINTIME_FAST ,
541.Dv BPF_T_NANOTIME_MONOTONIC ,
542or
543.Dv BPF_T_BINTIME_MONOTONIC_FAST
544to get time stamps in 64-bit
545.Vt struct bintime
546format.
547Set to
548.Dv BPF_T_NONE
549to ignore time stamp.
550All 64-bit time stamp formats are wrapped in
551.Vt struct bpf_ts .
552The
553.Dv BPF_T_MICROTIME_FAST ,
554.Dv BPF_T_NANOTIME_FAST ,
555.Dv BPF_T_BINTIME_FAST ,
556.Dv BPF_T_MICROTIME_MONOTONIC_FAST ,
557.Dv BPF_T_NANOTIME_MONOTONIC_FAST ,
558and
559.Dv BPF_T_BINTIME_MONOTONIC_FAST
560are analogs of corresponding formats without _FAST suffix but do not perform
561a full time counter query, so their accuracy is one timer tick.
562The
563.Dv BPF_T_MICROTIME_MONOTONIC ,
564.Dv BPF_T_NANOTIME_MONOTONIC ,
565.Dv BPF_T_BINTIME_MONOTONIC ,
566.Dv BPF_T_MICROTIME_MONOTONIC_FAST ,
567.Dv BPF_T_NANOTIME_MONOTONIC_FAST ,
568and
569.Dv BPF_T_BINTIME_MONOTONIC_FAST
570store the time elapsed since kernel boot.
571This setting is initialized to
572.Dv BPF_T_MICROTIME
573by default.
574.It Dv BIOCFEEDBACK
575.Pq Li u_int
576Set packet feedback mode.
577This allows injected packets to be fed back as input to the interface when
578output via the interface is successful.
579When
580.Dv BPF_D_INOUT
581direction is set, injected outgoing packet is not returned by BPF to avoid
582duplication. This flag is initialized to zero by default.
583.It Dv BIOCLOCK
584Set the locked flag on the
585.Nm
586descriptor.
587This prevents the execution of
588ioctl commands which could change the underlying operating parameters of
589the device.
590.It Dv BIOCGETBUFMODE
591.It Dv BIOCSETBUFMODE
592.Pq Li u_int
593Get or set the current
594.Nm
595buffering mode; possible values are
596.Dv BPF_BUFMODE_BUFFER ,
597buffered read mode, and
598.Dv BPF_BUFMODE_ZBUF ,
599zero-copy buffer mode.
600.It Dv BIOCSETZBUF
601.Pq Li struct bpf_zbuf
602Set the current zero-copy buffer locations; buffer locations may be
603set only once zero-copy buffer mode has been selected, and prior to attaching
604to an interface.
605Buffers must be of identical size, page-aligned, and an integer multiple of
606pages in size.
607The three fields
608.Vt bz_bufa ,
609.Vt bz_bufb ,
610and
611.Vt bz_buflen
612must be filled out.
613If buffers have already been set for this device, the ioctl will fail.
614.It Dv BIOCGETZMAX
615.Pq Li size_t
616Get the largest individual zero-copy buffer size allowed.
617As two buffers are used in zero-copy buffer mode, the limit (in practice) is
618twice the returned size.
619As zero-copy buffers consume kernel address space, conservative selection of
620buffer size is suggested, especially when there are multiple
621.Nm
622descriptors in use on 32-bit systems.
623.It Dv BIOCROTZBUF
624Force ownership of the next buffer to be assigned to userspace, if any data
625present in the buffer.
626If no data is present, the buffer will remain owned by the kernel.
627This allows consumers of zero-copy buffering to implement timeouts and
628retrieve partially filled buffers.
629In order to handle the case where no data is present in the buffer and
630therefore ownership is not assigned, the user process must check
631.Vt bzh_kernel_gen
632against
633.Vt bzh_user_gen .
634.El
635.Sh BPF HEADER
636One of the following structures is prepended to each packet returned by
637.Xr read 2
638or via a zero-copy buffer:
639.Bd -literal
640struct bpf_xhdr {
641	struct bpf_ts	bh_tstamp;     /* time stamp */
642	uint32_t	bh_caplen;     /* length of captured portion */
643	uint32_t	bh_datalen;    /* original length of packet */
644	u_short		bh_hdrlen;     /* length of bpf header (this struct
645					  plus alignment padding */
646};
647
648struct bpf_hdr {
649	struct timeval	bh_tstamp;     /* time stamp */
650	uint32_t	bh_caplen;     /* length of captured portion */
651	uint32_t	bh_datalen;    /* original length of packet */
652	u_short		bh_hdrlen;     /* length of bpf header (this struct
653					  plus alignment padding */
654};
655.Ed
656.Pp
657The fields, whose values are stored in host order, and are:
658.Pp
659.Bl -tag -compact -width bh_datalen
660.It Li bh_tstamp
661The time at which the packet was processed by the packet filter.
662.It Li bh_caplen
663The length of the captured portion of the packet.
664This is the minimum of
665the truncation amount specified by the filter and the length of the packet.
666.It Li bh_datalen
667The length of the packet off the wire.
668This value is independent of the truncation amount specified by the filter.
669.It Li bh_hdrlen
670The length of the
671.Nm
672header, which may not be equal to
673.\" XXX - not really a function call
674.Fn sizeof "struct bpf_xhdr"
675or
676.Fn sizeof "struct bpf_hdr" .
677.El
678.Pp
679The
680.Li bh_hdrlen
681field exists to account for
682padding between the header and the link level protocol.
683The purpose here is to guarantee proper alignment of the packet
684data structures, which is required on alignment sensitive
685architectures and improves performance on many other architectures.
686The packet filter ensures that the
687.Vt bpf_xhdr ,
688.Vt bpf_hdr
689and the network layer
690header will be word aligned.
691Currently,
692.Vt bpf_hdr
693is used when the time stamp is set to
694.Dv BPF_T_MICROTIME ,
695.Dv BPF_T_MICROTIME_FAST ,
696.Dv BPF_T_MICROTIME_MONOTONIC ,
697.Dv BPF_T_MICROTIME_MONOTONIC_FAST ,
698or
699.Dv BPF_T_NONE
700for backward compatibility reasons.  Otherwise,
701.Vt bpf_xhdr
702is used.  However,
703.Vt bpf_hdr
704may be deprecated in the near future.
705Suitable precautions
706must be taken when accessing the link layer protocol fields on alignment
707restricted machines.
708(This is not a problem on an Ethernet, since
709the type field is a short falling on an even offset,
710and the addresses are probably accessed in a bytewise fashion).
711.Pp
712Additionally, individual packets are padded so that each starts
713on a word boundary.
714This requires that an application
715has some knowledge of how to get from packet to packet.
716The macro
717.Dv BPF_WORDALIGN
718is defined in
719.In net/bpf.h
720to facilitate
721this process.
722It rounds up its argument to the nearest word aligned value (where a word is
723.Dv BPF_ALIGNMENT
724bytes wide).
725.Pp
726For example, if
727.Sq Li p
728points to the start of a packet, this expression
729will advance it to the next packet:
730.Dl p = (char *)p + BPF_WORDALIGN(p->bh_hdrlen + p->bh_caplen)
731.Pp
732For the alignment mechanisms to work properly, the
733buffer passed to
734.Xr read 2
735must itself be word aligned.
736The
737.Xr malloc 3
738function
739will always return an aligned buffer.
740.Sh FILTER MACHINE
741A filter program is an array of instructions, with all branches forwardly
742directed, terminated by a
743.Em return
744instruction.
745Each instruction performs some action on the pseudo-machine state,
746which consists of an accumulator, index register, scratch memory store,
747and implicit program counter.
748.Pp
749The following structure defines the instruction format:
750.Bd -literal
751struct bpf_insn {
752	u_short	code;
753	u_char 	jt;
754	u_char 	jf;
755	u_long k;
756};
757.Ed
758.Pp
759The
760.Li k
761field is used in different ways by different instructions,
762and the
763.Li jt
764and
765.Li jf
766fields are used as offsets
767by the branch instructions.
768The opcodes are encoded in a semi-hierarchical fashion.
769There are eight classes of instructions:
770.Dv BPF_LD ,
771.Dv BPF_LDX ,
772.Dv BPF_ST ,
773.Dv BPF_STX ,
774.Dv BPF_ALU ,
775.Dv BPF_JMP ,
776.Dv BPF_RET ,
777and
778.Dv BPF_MISC .
779Various other mode and
780operator bits are or'd into the class to give the actual instructions.
781The classes and modes are defined in
782.In net/bpf.h .
783.Pp
784Below are the semantics for each defined
785.Nm
786instruction.
787We use the convention that A is the accumulator, X is the index register,
788P[] packet data, and M[] scratch memory store.
789P[i:n] gives the data at byte offset
790.Dq i
791in the packet,
792interpreted as a word (n=4),
793unsigned halfword (n=2), or unsigned byte (n=1).
794M[i] gives the i'th word in the scratch memory store, which is only
795addressed in word units.
796The memory store is indexed from 0 to
797.Dv BPF_MEMWORDS
798- 1.
799.Li k ,
800.Li jt ,
801and
802.Li jf
803are the corresponding fields in the
804instruction definition.
805.Dq len
806refers to the length of the packet.
807.Pp
808.Bl -tag -width BPF_STXx
809.It Dv BPF_LD
810These instructions copy a value into the accumulator.
811The type of the source operand is specified by an
812.Dq addressing mode
813and can be a constant
814.Pq Dv BPF_IMM ,
815packet data at a fixed offset
816.Pq Dv BPF_ABS ,
817packet data at a variable offset
818.Pq Dv BPF_IND ,
819the packet length
820.Pq Dv BPF_LEN ,
821or a word in the scratch memory store
822.Pq Dv BPF_MEM .
823For
824.Dv BPF_IND
825and
826.Dv BPF_ABS ,
827the data size must be specified as a word
828.Pq Dv BPF_W ,
829halfword
830.Pq Dv BPF_H ,
831or byte
832.Pq Dv BPF_B .
833The semantics of all the recognized
834.Dv BPF_LD
835instructions follow.
836.Pp
837.Bd -literal
838BPF_LD+BPF_W+BPF_ABS	A <- P[k:4]
839BPF_LD+BPF_H+BPF_ABS	A <- P[k:2]
840BPF_LD+BPF_B+BPF_ABS	A <- P[k:1]
841BPF_LD+BPF_W+BPF_IND	A <- P[X+k:4]
842BPF_LD+BPF_H+BPF_IND	A <- P[X+k:2]
843BPF_LD+BPF_B+BPF_IND	A <- P[X+k:1]
844BPF_LD+BPF_W+BPF_LEN	A <- len
845BPF_LD+BPF_IMM		A <- k
846BPF_LD+BPF_MEM		A <- M[k]
847.Ed
848.It Dv BPF_LDX
849These instructions load a value into the index register.
850Note that
851the addressing modes are more restrictive than those of the accumulator loads,
852but they include
853.Dv BPF_MSH ,
854a hack for efficiently loading the IP header length.
855.Pp
856.Bd -literal
857BPF_LDX+BPF_W+BPF_IMM	X <- k
858BPF_LDX+BPF_W+BPF_MEM	X <- M[k]
859BPF_LDX+BPF_W+BPF_LEN	X <- len
860BPF_LDX+BPF_B+BPF_MSH	X <- 4*(P[k:1]&0xf)
861.Ed
862.It Dv BPF_ST
863This instruction stores the accumulator into the scratch memory.
864We do not need an addressing mode since there is only one possibility
865for the destination.
866.Pp
867.Bd -literal
868BPF_ST			M[k] <- A
869.Ed
870.It Dv BPF_STX
871This instruction stores the index register in the scratch memory store.
872.Pp
873.Bd -literal
874BPF_STX			M[k] <- X
875.Ed
876.It Dv BPF_ALU
877The alu instructions perform operations between the accumulator and
878index register or constant, and store the result back in the accumulator.
879For binary operations, a source mode is required
880.Dv ( BPF_K
881or
882.Dv BPF_X ) .
883.Pp
884.Bd -literal
885BPF_ALU+BPF_ADD+BPF_K	A <- A + k
886BPF_ALU+BPF_SUB+BPF_K	A <- A - k
887BPF_ALU+BPF_MUL+BPF_K	A <- A * k
888BPF_ALU+BPF_DIV+BPF_K	A <- A / k
889BPF_ALU+BPF_AND+BPF_K	A <- A & k
890BPF_ALU+BPF_OR+BPF_K	A <- A | k
891BPF_ALU+BPF_LSH+BPF_K	A <- A << k
892BPF_ALU+BPF_RSH+BPF_K	A <- A >> k
893BPF_ALU+BPF_ADD+BPF_X	A <- A + X
894BPF_ALU+BPF_SUB+BPF_X	A <- A - X
895BPF_ALU+BPF_MUL+BPF_X	A <- A * X
896BPF_ALU+BPF_DIV+BPF_X	A <- A / X
897BPF_ALU+BPF_AND+BPF_X	A <- A & X
898BPF_ALU+BPF_OR+BPF_X	A <- A | X
899BPF_ALU+BPF_LSH+BPF_X	A <- A << X
900BPF_ALU+BPF_RSH+BPF_X	A <- A >> X
901BPF_ALU+BPF_NEG		A <- -A
902.Ed
903.It Dv BPF_JMP
904The jump instructions alter flow of control.
905Conditional jumps
906compare the accumulator against a constant
907.Pq Dv BPF_K
908or the index register
909.Pq Dv BPF_X .
910If the result is true (or non-zero),
911the true branch is taken, otherwise the false branch is taken.
912Jump offsets are encoded in 8 bits so the longest jump is 256 instructions.
913However, the jump always
914.Pq Dv BPF_JA
915opcode uses the 32 bit
916.Li k
917field as the offset, allowing arbitrarily distant destinations.
918All conditionals use unsigned comparison conventions.
919.Pp
920.Bd -literal
921BPF_JMP+BPF_JA		pc += k
922BPF_JMP+BPF_JGT+BPF_K	pc += (A > k) ? jt : jf
923BPF_JMP+BPF_JGE+BPF_K	pc += (A >= k) ? jt : jf
924BPF_JMP+BPF_JEQ+BPF_K	pc += (A == k) ? jt : jf
925BPF_JMP+BPF_JSET+BPF_K	pc += (A & k) ? jt : jf
926BPF_JMP+BPF_JGT+BPF_X	pc += (A > X) ? jt : jf
927BPF_JMP+BPF_JGE+BPF_X	pc += (A >= X) ? jt : jf
928BPF_JMP+BPF_JEQ+BPF_X	pc += (A == X) ? jt : jf
929BPF_JMP+BPF_JSET+BPF_X	pc += (A & X) ? jt : jf
930.Ed
931.It Dv BPF_RET
932The return instructions terminate the filter program and specify the amount
933of packet to accept (i.e., they return the truncation amount).
934A return value of zero indicates that the packet should be ignored.
935The return value is either a constant
936.Pq Dv BPF_K
937or the accumulator
938.Pq Dv BPF_A .
939.Pp
940.Bd -literal
941BPF_RET+BPF_A		accept A bytes
942BPF_RET+BPF_K		accept k bytes
943.Ed
944.It Dv BPF_MISC
945The miscellaneous category was created for anything that does not
946fit into the above classes, and for any new instructions that might need to
947be added.
948Currently, these are the register transfer instructions
949that copy the index register to the accumulator or vice versa.
950.Pp
951.Bd -literal
952BPF_MISC+BPF_TAX	X <- A
953BPF_MISC+BPF_TXA	A <- X
954.Ed
955.El
956.Pp
957The
958.Nm
959interface provides the following macros to facilitate
960array initializers:
961.Fn BPF_STMT opcode operand
962and
963.Fn BPF_JUMP opcode operand true_offset false_offset .
964.Sh FILES
965.Bl -tag -compact -width /dev/bpf
966.It Pa /dev/bpf
967the packet filter device
968.El
969.Sh EXAMPLES
970The following filter is taken from the Reverse ARP Daemon.
971It accepts only Reverse ARP requests.
972.Bd -literal
973struct bpf_insn insns[] = {
974	BPF_STMT(BPF_LD+BPF_H+BPF_ABS, 12),
975	BPF_JUMP(BPF_JMP+BPF_JEQ+BPF_K, ETHERTYPE_REVARP, 0, 3),
976	BPF_STMT(BPF_LD+BPF_H+BPF_ABS, 20),
977	BPF_JUMP(BPF_JMP+BPF_JEQ+BPF_K, REVARP_REQUEST, 0, 1),
978	BPF_STMT(BPF_RET+BPF_K, sizeof(struct ether_arp) +
979		 sizeof(struct ether_header)),
980	BPF_STMT(BPF_RET+BPF_K, 0),
981};
982.Ed
983.Pp
984This filter accepts only IP packets between host 128.3.112.15 and
985128.3.112.35.
986.Bd -literal
987struct bpf_insn insns[] = {
988	BPF_STMT(BPF_LD+BPF_H+BPF_ABS, 12),
989	BPF_JUMP(BPF_JMP+BPF_JEQ+BPF_K, ETHERTYPE_IP, 0, 8),
990	BPF_STMT(BPF_LD+BPF_W+BPF_ABS, 26),
991	BPF_JUMP(BPF_JMP+BPF_JEQ+BPF_K, 0x8003700f, 0, 2),
992	BPF_STMT(BPF_LD+BPF_W+BPF_ABS, 30),
993	BPF_JUMP(BPF_JMP+BPF_JEQ+BPF_K, 0x80037023, 3, 4),
994	BPF_JUMP(BPF_JMP+BPF_JEQ+BPF_K, 0x80037023, 0, 3),
995	BPF_STMT(BPF_LD+BPF_W+BPF_ABS, 30),
996	BPF_JUMP(BPF_JMP+BPF_JEQ+BPF_K, 0x8003700f, 0, 1),
997	BPF_STMT(BPF_RET+BPF_K, (u_int)-1),
998	BPF_STMT(BPF_RET+BPF_K, 0),
999};
1000.Ed
1001.Pp
1002Finally, this filter returns only TCP finger packets.
1003We must parse the IP header to reach the TCP header.
1004The
1005.Dv BPF_JSET
1006instruction
1007checks that the IP fragment offset is 0 so we are sure
1008that we have a TCP header.
1009.Bd -literal
1010struct bpf_insn insns[] = {
1011	BPF_STMT(BPF_LD+BPF_H+BPF_ABS, 12),
1012	BPF_JUMP(BPF_JMP+BPF_JEQ+BPF_K, ETHERTYPE_IP, 0, 10),
1013	BPF_STMT(BPF_LD+BPF_B+BPF_ABS, 23),
1014	BPF_JUMP(BPF_JMP+BPF_JEQ+BPF_K, IPPROTO_TCP, 0, 8),
1015	BPF_STMT(BPF_LD+BPF_H+BPF_ABS, 20),
1016	BPF_JUMP(BPF_JMP+BPF_JSET+BPF_K, 0x1fff, 6, 0),
1017	BPF_STMT(BPF_LDX+BPF_B+BPF_MSH, 14),
1018	BPF_STMT(BPF_LD+BPF_H+BPF_IND, 14),
1019	BPF_JUMP(BPF_JMP+BPF_JEQ+BPF_K, 79, 2, 0),
1020	BPF_STMT(BPF_LD+BPF_H+BPF_IND, 16),
1021	BPF_JUMP(BPF_JMP+BPF_JEQ+BPF_K, 79, 0, 1),
1022	BPF_STMT(BPF_RET+BPF_K, (u_int)-1),
1023	BPF_STMT(BPF_RET+BPF_K, 0),
1024};
1025.Ed
1026.Sh SEE ALSO
1027.Xr tcpdump 1 ,
1028.Xr ioctl 2 ,
1029.Xr kqueue 2 ,
1030.Xr poll 2 ,
1031.Xr select 2 ,
1032.Xr byteorder 3 ,
1033.Xr ng_bpf 4 ,
1034.Xr bpf 9
1035.Rs
1036.%A McCanne, S.
1037.%A Jacobson V.
1038.%T "An efficient, extensible, and portable network monitor"
1039.Re
1040.Sh HISTORY
1041The Enet packet filter was created in 1980 by Mike Accetta and
1042Rick Rashid at Carnegie-Mellon University.
1043Jeffrey Mogul, at
1044Stanford, ported the code to
1045.Bx
1046and continued its development from
10471983 on.
1048Since then, it has evolved into the Ultrix Packet Filter at
1049.Tn DEC ,
1050a
1051.Tn STREAMS
1052.Tn NIT
1053module under
1054.Tn SunOS 4.1 ,
1055and
1056.Tn BPF .
1057.Sh AUTHORS
1058.An -nosplit
1059.An Steven McCanne ,
1060of Lawrence Berkeley Laboratory, implemented BPF in
1061Summer 1990.
1062Much of the design is due to
1063.An Van Jacobson .
1064.Pp
1065Support for zero-copy buffers was added by
1066.An Robert N. M. Watson
1067under contract to Seccuris Inc.
1068.Sh BUGS
1069The read buffer must be of a fixed size (returned by the
1070.Dv BIOCGBLEN
1071ioctl).
1072.Pp
1073A file that does not request promiscuous mode may receive promiscuously
1074received packets as a side effect of another file requesting this
1075mode on the same hardware interface.
1076This could be fixed in the kernel with additional processing overhead.
1077However, we favor the model where
1078all files must assume that the interface is promiscuous, and if
1079so desired, must utilize a filter to reject foreign packets.
1080.Pp
1081Data link protocols with variable length headers are not currently supported.
1082.Pp
1083The
1084.Dv SEESENT ,
1085.Dv DIRECTION ,
1086and
1087.Dv FEEDBACK
1088settings have been observed to work incorrectly on some interface
1089types, including those with hardware loopback rather than software loopback,
1090and point-to-point interfaces.
1091They appear to function correctly on a
1092broad range of Ethernet-style interfaces.
1093