Re: [PATCH 2/2] Wait for child on signal death for aliases to externals

On Sun, Jul 05, 2020 at 12:18:38AM +0200, trygveaa@gmail.com wrote:

Usually when I find myself saying "the previous commit" and "does the same" in a commit message, it's a good indication that it might be worth combining the two commits into one. But in this case I think you're right to keep them separate; the other one was a regression fix, and this would be a change in behavior.

But it might be worth explaining more what the user-visible behavior we're trying to solve is.

Let's try to break down what differences the user might see.

The original is using the full run_command(), so we'd always be waiting for the alias command to finish before exiting under normal circumstances. The only difference is that now when die due to a signal, we'd wait then, too.

And that only matters if:

  - we've spawned a pager that doesn't die due to the signal, and we
    want to wait until it has finished. However, that seems to work
    already for me in a quick test of:

      $ git -p -c alias.foo='!echo foo; sleep 10; echo bar' foo
      ^C

    where the pager keeps running (because it ignores the signal), and
    the main git process doesn't exit either (because it's waiting for
    the pager to finish)

    I think this case is different than the one fixed by 46df6906f3
    (execv_dashed_external: wait for child on signal death, 2017-01-06)
    because the pager is spawned by the main Git process (so it knows to
    wait), rather than the child dashed-external.

    I guess to recreate that you'd need to trigger the pager inside the
    alias itself, like:

      $ git -c alias.foo='!{ echo foo; sleep 10; echo bar; } | less' foo
      ^C

    which does exhibit the annoying behavior (we exit, and pgrp loses
    the tty session leader bit, and the pager gets EIO).

  - the child for some reason decides not to respect the signal.
    Obviously running a pager is one reason it would decide to do so,
    and we'd be improving the behavior there. I have trouble imagining a
    case where waiting for it would do the _wrong_ thing. I.e., if I do:

      $ git -c alias.foo='!trap "echo got signal" INT; sleep 5' foo

    it probably does make sense for us to continue to wait on that alias
    to complete (I'm not sure what anyone would try to accomplish with
    that, but waiting seems like the least-astonishing thing to me).

So I think this is moving in the right direction.

However, there is one weirdness worth thinking about. Because the wait_after_clean feature relies on actually trying to clean the child, Git will actually send a signal to the alias shell. For a signal, we'll try to pass along the same signal, so it just means the shell will get SIGINT twice in this case (or more likely, they'll race and we'll ignore the duplicate signal while the child is in its own handler).

We'd likewise send a signal if we hit a normal exit(), but we shouldn't do so because we're already blocked waiting for the child to exit.

But a more interesting case is if somebody sends the parent git process a signal but _not_ the child (e.g., kill -TERM). Then we'd pass SIGTERM along to the child. I'd venture to say that's _probably_ the right thing to do, if only because it's exactly what we do for a dashed external as well.

Sorry that ended up so long-winded, but my conclusion is: this is doing the right thing. We should probably embed some of that discussion in the commit message, but I think the simplest argument is just: this is what we do for external commands we run, so treating shell aliases the same in terms of passing along signals makes things simple and consistent.

-Peff