Move pseudoterminal slave open to child

Hopefully this fixes "unexpected EOF" failures on macOS
(#3137, #3605, #7242, #7702).

The problem appears to be that under some circumstances, macOS
discards the output written to the slave side of the
pseudoterminal. Hence the parent never sees the "sandbox initialized"
message from the child, even though it succeeded. The conditions are:

* The child finishes very quickly. That's why this bug is likely to
  trigger in nix-env tests, since that uses a builtin builder. Adding
  a short sleep before the child exits makes the problem go away.

* The parent has closed its duplicate of the slave file
  descriptor. This shouldn't matter, since the child has a duplicate
  as well, but it does. E.g. moving the close to the bottom of
  startBuilder() makes the problem go away. However, that's not a
  solution because it would make Nix hang if the child dies before
  sending the "sandbox initialized" message.

* The system is under high load. E.g. "make installcheck -j16" makes
  the issue pretty reproducible, while it's very rare under "make
  installcheck -j1".

As a fix/workaround, we now open the pseudoterminal slave in the
child, rather than the parent. This removes the second condition
(i.e. the parent no longer needs to close the slave fd) and I haven't
been able to reproduce the "unexpected EOF" with this.

1 files changed, 4 insertions, 0 deletions

diff --git a/src/libstore/build/local-derivation-goal.hh b/src/libstore/build/local-derivation-goal.hh
index 34c4e9187..c8537c7e5 100644
--- a/src/libstore/build/local-derivation-goal.hh
+++ b/src/libstore/build/local-derivation-goal.hh
@@ -27,6 +27,10 @@ struct LocalDerivationGoal : public DerivationGoal
     /* Pipe for the builder's standard output/error. */
     Pipe builderOut;
 
+    /* Slave side of the pseudoterminal used for the builder's
+       standard output/error. */
+    Path slaveName;
+
     /* Pipe for synchronising updates to the builder namespaces. */
     Pipe userNamespaceSync;