[John Thacker <johnthacker@xxxxxxxxx>]

On Mon, Jun 21, 2021 at 9:28 PM João Valverde via Wireshark-dev <wireshark-dev@xxxxxxxxxxxxx> wrote:

On 22/06/21 01:26, John Thacker wrote:
> On Mon, Jun 21, 2021 at 2:21 PM João Valverde via Wireshark-dev
> <wireshark-dev@xxxxxxxxxxxxx <mailto:wireshark-dev@xxxxxxxxxxxxx>> wrote:
>
>
>
>     On 16/06/21 15:36, David Perry wrote:
>      > Sorry to drag up an old topic, but I've been thinking about this:
>      >
>      >> Message: 5
>      >> Date: Sat, 29 May 2021 09:32:29 +0200
>      >> From: Anders Broman <a.broman58@xxxxxxxxx
>     <mailto:a.broman58@xxxxxxxxx>>
>      >> To: Developer support list for Wireshark
>     <wireshark-dev@xxxxxxxxxxxxx <mailto:wireshark-dev@xxxxxxxxxxxxx>>
>      >> Subject: Re: [Wireshark-dev] Calling a dissector: Type for data
>      >>     parameter
>      >> Message-ID:
>      >>
>          <CAOpyz=zDycm33PXUwtBCTew7gTTEcSLiJ-f8SHW0L-863Q517A@xxxxxxxxxxxxxx <mailto:zDycm33PXUwtBCTew7gTTEcSLiJ-f8SHW0L-863Q517A@xxxxxxxxxxxxxx>>
>      >> Content-Type: text/plain; charset="utf-8"
>      >>
>      >> Hi,
>      >> Yes the method is fragile. At the time of development I think it was
>      >> proposed to pass a struct containing a string and the void
>     pointer where
>      >> the string could be used as a identifier. But that was voted down.
>      >> Regards
>      >> Anders
>      >
>      > I wasn't around for that discussion so I don't know the reasons, but
>      > how does this sound as a refined approach?:
>      >
>      > * Define a `dissector_data_t` that has a `guint32` identifier field,
>      > and a `void *` data field.
>      >
>      > * Replace the `void *data` parameter to dissectors with a pointer
>     to a
>      > `dissector_data_t`.
>      >
>      > * Either:
>      >
>      >     * Easy way: maintain a static list of identifiers that map to
>      > expected data types, or
>      >
>      >     * Have dissector X request an identifier in its registration
>      > function for the type of data it expects, and have dissector Y
>     (which
>      > will call X) request, in its handoff function, the identifier of the
>      > type of data it needs to pass to X.
>      >
>      > * Dissectors check for the right identifier in their
>      > `dissector_data_t` parameter and don't try to use it if it's wrong.
>      >
>      > Thoughts?
>      >
>
>     I think what you suggest would be the most straightforward fix.
>
>     To avoid breaking backward compatibility and changing thousands of
>     dissectors at the same time, both of which are highly problematic, it
>     can be done by adding a new dissector type (like it was done with
>     "dissector_cb_t", only using a different signature).[1]
>
>     Also a giant static list of dissector_data_t type identifiers would be
>     pretty clunky. I think we could recycle the protocol registration
>     number
>     for that.
>
>
> Perhaps I don't quite understand, but what would be the point if the
> protocol registration number were used? Presumably that is the number
> for the called protocol, based on what David outlined (the called
> protocol registering what data it expects.) But the calling dissector
> would always have that number (via dissector_handle_get_protocol_index()
> ) and pass it in, which wouldn't provide any guarantee that the data
> passed in was the correct type than what is being done now.
>
> The only way that I see it would make sense to pass in an identifier is
> if a protocol registers multiple data types it might expect to be passed
> in when called from different types (whether in one dissect_proto()
> function or multiple ones), in which case the protocol registration
> number couldn't be used, or if the identifier is instead related to the
> calling protocol and controlled by it (which is perhaps for this method
> of calling the wrong dependency direction, unlike with dissector tables
> where the calling protocol does control the passed data type, e.g.
> packet-ip always passes a ws_ip4* to the "ip proto" table or its
> heuristic subdissector table.)
>
> That doesn't sound like what's being proposed, though, so I am confused.

The way I proposed it the identifier is related to the calling protocol
and controlled by it. It refers to the first suggestion above: "Easy
way: maintain a static list of identifiers that map to expected data types"

Instead of having the biggest enum ever of static data identifiers we
can use the protocol registration number. There would probably be other
fields too, like version and flags, TBD).

Dissectors registering on the "ip.proto" table can receive a ws_ip4 or a
ws_ip6, the first byte is the type, so I don't follow your example
either. It's not really true that there is always a ws_ip4 pointer.

Ah, yes, I see, packet-ipv6 calls the same ip_try_dissect() function from packet-ip, only with a ws_ip6 as the void* iph parameter. So I had that quite wrong, and that's a good case for why the current situation is fragile, since any function called in that table (or the heuristic table) can be passed either a ws_ip4 or ws_ip6 pointer. Though at least as you say they can look at the first byte to determine the version (as in ws_ip_protocol() and packet-tapa.c). Though in practice most of them don't do anything with the passed in data, aside from ICMP, ICMP for IPv6, and the IPv6 extension header types that can reasonably guarantee what they're getting.

The way I read the previous suggestion you responded to ("dissector X request an identifier in its registration function for the type of data it expects") made it sound like the called function would register the type of data it expected, in which case the "list of static data identifiers" would have been indexed by the called protocol, or by some reduced set if some protocols expected the same data. My assumption is that the motivation was for the called protocols to be able to determine if the data passed in from a calling protocol (possibly a newly written dissector or plugin) was the correct type without having to know specifically about the calling protocol, even in the case of the dissector being a newly written plugin with a previously unknown protocol registration id. I suppose that could still be possible that way.

John