Description of problem:
common/asshelp.c:start_new_gpg_agent does (skipping error handling and similar non-important code):
>          err = gnupg_spawn_process_detached (program? program : agent_program,
>                                              argv, NULL);
>              for (i=0; i < SECS_TO_WAIT_FOR_AGENT; i++)
>                {
>                  gnupg_sleep (1);
>                  err = assuan_socket_connect (ctx, sockname, 0, 0);
>                  if (!err)
>                      break;
>                }

As a result, even when the agent is ready to accept a connection very quickly (less than 2 ms in my case), gpg2 is waiting for a _whole second_ before trying to connect at all.

This extremely noticeably slows down automated gpg runs, where each run creates a new $GNUPGHOME and imports a set of keys into it.

Please speed this up, perhaps implementing exponential backoff (1/2/4/8/… ms up to the current 1 s, or so), or perhaps teaching gpg-agent to close a single-purpose pipe after its listen() succeeds, so that gpg-agent can create such a p ipe and block until it is closed = try to connect as soon as possible instead of having to guess how long to wait.

Version-Release number of selected component (if applicable):
gnupg2-2.1.13-2.fc24.x86_64 (observation)
gnupg2-2.1.18-2.fc26.src.rpm (examined source code)


Steps to Reproduce:
1. mkdir t
2. (time GNUPGHOME=$(pwd)/t gpg2 --list-keys)
3. (time GNUPGHOME=$(pwd)/t gpg2 --import < ~/.gnupg/pubring.gpg)
3. (time GNUPGHOME=$(pwd)/t gpg2 --import < ~/.gnupg/pubring.gpg)

Actual results:
2. takes 0.004 seconds (all values real time)
3. autostarts the agent, and takes 1.012 seconds
4. reuses an existing agent, and takes 0.013 seconds

Expected results:
3. takes less than, say, 0.30 seconds