Re: [WIP RFC 4/5] upload-pack: refactor writing of "packfile" line

From: Junio C Hamano <gitster@pobox.com>
Date: Dec 7, 2018, 00:22 UTC
Message-ID: <xmqqmupi89ub.fsf@gitster-ct.c.googlers.com>
In-Reply-To: <20181206232538.141378-1-jonathantanmy@google.com>

Jonathan Tan <jonathantanmy@google.com> writes:

Show 20 quoted lines

>> Jonathan Tan <jonathantanmy@google.com> writes:
>> 
>> > @@ -126,6 +129,12 @@ static int read_pack_objects_stdout(int outfd, struct output_state *os)
>> >  	}
>> >  	os->used += readsz;
>> >  
>> > +	if (!os->packfile_started) {
>> > +		os->packfile_started = 1;
>> > +		if (use_protocol_v2)
>> > +			packet_write_fmt(1, "packfile\n");
>> 
>> If we fix this function so that the only byte in the buffer is held
>> back without emitted when os->used == 1 as I alluded to, this may
>> have to be done a bit later, as with such a change, it is no longer
>> guaranteed that send_client_data() will be called after this point.
>
> I'm not sure what you mean about there being no guarantee that
> send_client_data() is not called - in create_pack_file(), there is an
> "if (output_state.used > 0)" line (previously "if (0 <= buffered)") that
> outputs anything remaining.

I was referring to this part of the review on the previous step, which you may not yet have read.

    OK, this corresponds to the "*cp++ = buffered"  in the original just
    before xread().

    > +		os->used = 1;
    > +	} else {
    > +		send_client_data(1, os->buffer, os->used);
    > +		os->used = 0;

    I am not sure if the code is correct when os->used happens to be 1
    (shouldn't we hold the byte, skip the call to send_client_data(),
    and go back to poll() to expect more data?), but this is a faithful
    code movement and rewrite of the original.

The point of this logic is to make sure we always hold back some bytes and do not feed *all* the bytes to the other side by calling "send-client-data" until we made sure the upstream of what we are relaying (pack-objects?) successfully exited, but it looks to me that the "else" clause above ends up flushing everything when os->used is 1, which goes against the whole purpose of the code.

And the "fix" I was alluding to was to update that "else" clause to make it a no-op that keeps os->used non-zero, which would not call send-client-data.

When that fix happens, the part that early in the function this patch added "now we know we will call send-client-data, so let's say 'here comes packdata' unless we have already said that" is making the decision too early. Depending on the value of os->used when we enter the code and the number of bytes xread() reads from the upstream, we might not call send-client-data yet (namely, when we have no buffered data and we happened to get only one byte).

> ... it might be
> better if the server can send sideband throughout the whole response -
> perhaps that should be investigated first.

Yup. It just looked quite crazy, and it is even more crazy to buffer keepalives ;-)