Re: [PATCH 2/2] send-pack: check atomic push before running GPG

From: Junio C Hamano <gitster@pobox.com>
Date: Sep 16, 2020, 04:37 UTC
Message-ID: <xmqqeen2xrok.fsf@gitster.c.googlers.com>
In-Reply-To: <CANYiYbHYi70ZcjDTyQ++_+njuZMF=TksPepH+wP+zNmhBABNAg@mail.gmail.com>

Jiang Xin <worldhello.net@gmail.com> writes:

Show 26 quoted lines

>> > -
>> > -     if (!args->dry_run && push_cert_nonce)
>> > -             cmds_sent = generate_push_cert(&req_buf, remote_refs, args,
>> > -                                            cap_buf.buf, push_cert_nonce);
>> > -
>> >       /*
>> >        * Clear the status for each ref and see if we need to send
>> >        * the pack data.
>>
>> This "Clear the status for each ref" worries me.
>>
>> The generate_push_cert() function RELIES on ref->status and filters
>> out the ref that failed to pass the local check from the generated
>> push certificate.  If you let the loop (post context of this hunk)
>> run, ref->status will be updated by it, so the net effect of this
>> patch is that it breaks "non-atomic" case that pushes multiple refs
>> and one of ref fails to pass the local check.
>>
>> IOW, generate_push_cert() MUST be called before this loop "clears
>> the status for each ref" by assigning to ref->status.
>
> The next block ("Finally, tell the other end!") is what we send
> commands to "receive-pack", right after some of the status are reset
> to REF_STATUS_OK or REF_STATUS_EXPECTING_REPORT by this chunk of code.
> So moving the generate_push_cert() part right before the "Finally,
> tell the other end!" part LGTM.

Sorry, I do not follow. The loop in question is the one before "Finally tell the other end". The loop ends like so:

	for (ref = remote_refs; ref; ref = ref->next) {
		...
		if (args->dry_run || !status_report)
			ref->status = REF_STATUS_OK;
		else
			ref->status = REF_STATUS_EXPECTING_REPORT;
	}

and the patch moves a call to generate_push_cert() that looks at remote_refs _after_ this loop, but generate_push_cert() function uses a loop over remote_refs that uses check_to_send_update(), which looks at ref->status's value to decide what to do. Its correct operation relies on ref->status NOT updated by the above loop.

The loop prepares the status field so that we can then read and record the response against each ref updates from the other side.

The ref->status field is set to EXPECTING_REPORT, later to be updated to REF_STATUS_OK or REF_STATUS_REMOTE_REJECT. We can clobber the original value of ref->status at this point only because the loop depends on the fact that no check_to_send_update() call will happen after the loop (because the ref->status field the helper's operation depends on is already reset for the next phase of operation). The patch that moves generate_push_cert() call below the loop, whether it is before or after the "Finally tell the other end" loop, is therefore fundamentally broken, isn't it?

I do not think it would introduce such breakage if we teach generate_push_cert() to pay attention to the atomicity, and that can be done without reordering the calls in send_pack() to break the control flow.