Re: [PATCH v2 2/2] connect, protocol: log negotiated protocol version

From: "Ævar Arnfjörð Bjarmason" <avarab@gmail.com>
To: Taylor Blau <me@ttaylorr.com>
Cc: Josh Steadmon <steadmon@google.com>,
	git@vger.kernel.org, gitster@pobox.com
Subject: Re: [PATCH v2 2/2] connect, protocol: log negotiated protocol version
Date: Thu, 05 Aug 2021 04:47:16 +0200	[thread overview]
Message-ID: <874kc4mwqc.fsf@evledraar.gmail.com> (raw)
In-Reply-To: <YQs+VQIYDO3pkCNS@nand.local>

On Wed, Aug 04 2021, Taylor Blau wrote:

> On Thu, Aug 05, 2021 at 01:40:51AM +0200, Ævar Arnfjörð Bjarmason wrote:
>>
>> On Wed, Aug 04 2021, Josh Steadmon wrote:
>>
>> > It is useful for performance monitoring and debugging purposes to know
>> > the wire protocol used for remote operations. This may differ from the
>> > version set in local configuration due to differences in version and/or
>> > configuration between the server and the client. Therefore, log the
>> > negotiated wire protocol version via trace2, for both clients and
>> > servers.
>> >
>> > Signed-off-by: Josh Steadmon <steadmon@google.com>
>> > ---
>>
>> I know Taylor asked you to change it to a string from in int upthread in
>> <YQmxSxTswHE/gTet@nand.local>, but I really don't see the point. But am
>> willing to be convinced otherwise.
>
> The conversion to log a string instead of an integer is necessary if
> Josh wants to write "<unknown>" instead of -1. To me, that seemed
> clearer, and I like that it makes the trace2 representation for a
> protocol version separate from the protocol_version enum.

Yes, having a magic -1 value would be bad, but since it seems we'll
never get it in practice...

>> It seems to me that both of these codepaths will never usefully use this
>> new "UNKNOWN_VERSION" string you added, i.e.:
>>
>> >  connect.c                             |  3 +++
>> >  protocol.c                            |  3 +++
>> >  t/t5705-session-id-in-capabilities.sh | 12 ++++++++++++
>> >  3 files changed, 18 insertions(+)
>> >
>> > diff --git a/connect.c b/connect.c
>> > index 70b13389ba..5f0e113625 100644
>> > --- a/connect.c
>> > +++ b/connect.c
>> > @@ -150,6 +150,9 @@ enum protocol_version discover_version(struct packet_reader *reader)
>> >  		break;
>> >  	}
>> >
>> > +	trace2_data_string("transfer", NULL, "negotiated-version",
>> > +			   format_protocol_version(version));
>>
>> Right after this.
>>
>> >  	switch (version) {
>> >  	case protocol_v2:
>> >  		process_capabilities_v2(reader);
>>
>> We'll die here with BUG("unknown protocol version") if it's unknown..
>
> Good eyes. In fact, the second switch statement shouldn't even need a
> case-arm for protocol_unknown_version (but has it to cover all
> enumerated values).

I didn't check if crafting an unknown version will be found earlier, or
if we'll actually reach that "unknown" case.

> I didn't realize before that the unknown case really is dead code, so
> we'll never log "<unknown>". And since the mapping from protocol_version
> to string is identical for known values, we could probably do without
> it.
>
> And I don't much care either way. I think the benefit is really pretty
> slim, and arguably my code is just adding unnecessary overhead. So I'm
> happy to go with or without it, but I'd be rather sad to spend much more
> of our collective time discussing it.

Yeah, I just think if we can be sure it's an integer *and* a valid
version when we log it, people writing future log summarizing code will
thank us, i.e. just 0, 1, 2, and in the future maybe 3, ..., but not -1
or "<unknown>" or other values we'll trust die() etc. to handle.