Re: [PATCH v2] receive-pack: add option to interrupt pre-receive when client exits

From: Robin Jarry <robin.jarry@6wind.com>
Date: Jan 27, 2022, 20:53 UTC
Message-ID: <CHGR6XNP6TV7.15VGVNQUJM9J6@diabtop>
In-Reply-To: <xmqqr18t2fxl.fsf@gitster.g>

Junio C Hamano, Jan 27, 2022 at 19:26:

Show 7 quoted lines

> Sorry, but I was (and am) questioning why we want to do more than
> "let it be killed by SIGPIPE, just like we used to do before
> ec7dbd14 (receive-pack: allow hooks to ignore its standard input
> stream, 2014-09-12) introduced the current behaviour", so the answer
> is still "why do we even need to complicate the thing with keepalive
> or anything we don't have, and we didn't have before ec7dbd14, in
> the code paths that are involved?"

My main goal is to abort a push if a user hits ctrl-c (or is disconnected) before the objects have been moved to permanent storage.

(partially) reverting to previous behavior would only allow aborting pushes *if* the pre-receive hook sends some output and this output cannot be forwarded to the client. There is no guarantee that the hook will send any output. Also, it would restore hard to track issue with poorly written pre-receive hooks.

I wonder if it would be possible to not rely on the pre-receive hook sending output and this output somehow not being forwarded to the client.

Instead, explicitly check if the client is still connected and alive after the pre-receive hook has exited but before completing the push transaction. That was my intent with that (invalid by the way) keepalive example.

I do not know git internals to say if it feasible without any protocol breakage. My attempts work well for aborting pushes:

 Writing objects: 100% (3/3), 321 bytes | 321.00 KiB/s, done.
 Total 3 (delta 1), reused 1 (delta 0), pack-reused 0
 remote: pre-receive start^C
 <--- client has disconnected
      receive-pack fails to send the "keepalive" packet
      the temp objects are *not* migrated to permanent storage

But this always leads to errors on the client side when receive-pack sends the "keepalive packet":

 Writing objects: 100% (3/3), 321 bytes | 321.00 KiB/s, done.
 Total 3 (delta 1), reused 1 (delta 0), pack-reused 0
 remote: pre-receive start
 remote: pre-receive end OK
 error: unexpected flush packet while reading remote unpack status
 error: invalid ref status from remote: unpack
 remote: post-receive start
 remote: post-receive end OK
 To git@host:repo.git
  ! [remote failure]    main -> main (remote failed to report status)
 error: failed to push some refs to 'git@host:repo.git'

Am I chasing rainbows or is that possible in the current state of the git protocol? Maybe I need to send the keepalive packet in a sideband? I have read the technical docs several times but I cannot understand how everything works properly.

Thank you for your time.