[tcpdump-workers] endianness of portable BPF bytecode

Denis Ovsienko denis at ovsienko.info
Fri Jun 10 18:37:41 EDT 2022


On Fri, 10 Jun 2022 14:26:34 -0700
Guy Harris <gharris at sonic.net> wrote:

> On Jun 10, 2022, at 1:59 PM, Denis Ovsienko via tcpdump-workers
> <tcpdump-workers at lists.tcpdump.org> wrote:
> 
> > Below is a draft of such a file format.  It addresses the following
> > needs:
> > * There is a header with a signature string to avoid false positive
> >  detection as some other file type that begins exactly with
> > particular bytecode (ran into this during disassembly experiments).
> > * There are version fields to address possible future changes to the
> >  encoding (either backward-compatible or not).  
> 
> Is the idea that a change that's backward-compatible (so that code
> that handles the new format needs no changes to handle the old
> format, but code that handles only the old format can't handle the
> new format) would involve a change to the minor version number, but a
> change that's not backward-compatible (so that to handle both
> versions would require two code paths for the two versions) would
> involve a change to the major version number?

Yes, more or less.  The draft format had a couple more fields not long
ago, with those the version semantics seemed more apparent (for a while
I thought Linux kernel cBPF is a superset of libpcap cBPF, but upon a
closer inspection of the header files they seem identical).

In any case, it is very convenient to be able to cycle a major version
and to redefine everything beyond the signature and the version fields,
that's the idea. Forward- and backward-compatibility between minor
versions can be considered now.

> > File format:
> > 
> > 0                   1                   2                   3
> > 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1
> > +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
> > |      'c'      |      'B'      |      'P'      |      'F'      |
> > +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+  
> 
> Is the 'c' part of the retronym "cBPF" for the "classic BPF"
> instruction set, as opposed to the eBPF instruction set?  (I didn't
> find any file format for saving eBPF programs, so this format could
> be used for that as well, with the magic number 'e' 'B' 'P' 'F'.)

Yes, it is.  In online documentation "eBPF" seems to clash with "BPF"
a lot, so it seems better to avoid the confusion early.

As it turned out after some research, the nominal binary format for eBPF
is ELF.  This is one of the most useful online documents I found:
https://www.man7.org/linux/man-pages/man8/tc-bpf.8.html
As you can see there and in the references into Linux kernel
documentation, ELF eBPF seems to cover different bit widths, relocation
types, debug information, lookup maps, multiple executable sections and
what not.

However, most of these features significantly overshoot the packet
capture problem space on one hand, and don't seem to address simple
practical needs of capturing parameters and context of a cBPF
compilation and reproducing it later.  So I figured it would be better
to leave eBPF solution space alone and to use a separate
purpose-designed file format for cBPF.

Most of the meta-data TLVs below are purposed to help a developer to
understand the context and reproduce the compilation.

> > Type=0x02 (LINKTYPE_ID)
> > Length=4
> > Value=<integer, link-layer header type>  
> 
> This could be 2 bytes long - pcapng limits link-layer types to 16
> bits, and pcap now can use the upper 16 bits of the link-layer type
> field for other purposes.

Fine.

> > Type=0x03 (LINKTYPE_NAME)
> > Length is variable
> > Value=<ASCII string, the link-layer header type name>  
> 
> E.g. either its LINKTYPE_xxx name or its DLT_xxx name?

Yes.  The intent is to capture the input to
pcap_datalink_name_to_val() if the latter was involved.

> > Type=0x04 (COMMENT)
> > Length is variabe
> > Value=<UTF-8 string, comment or the generating software
> > description>  
> 
> "Generating software description" as in the code that generated the
> BPF program?

"libpcap x.y.z", "my script v1.0" or something like that.

> > Type=0x05 (TIMESTAMP)
> > Length=8
> > Value=<integer, Unix timestamp>  
> 

> Is this the time the code was generated?

Yes.

> Is it a 64-bit time_t, or a 32-bit time_t and a 32-bit
> microseconds/nanoseconds value?  I'd recommend the former, unless we
> expect classic BPF to be dead by 2038.

It is the 64-bit integer time.

-- 
    Denis Ovsienko


More information about the tcpdump-workers mailing list