libstore: use vfork for linux sandbox launch wrappers

this halves sandbox launch overhead, reducing the build time for 3000
trivial runCommand derivations on our machine from 80 seconds to 44s.
as a nice side effect we also get better error message propagation in
some cases, most notably setgroups failing when run in lix sandboxes.

Change-Id: Ia7c50a844915d9b8a20475b90b2d0179fd2fff34

doc/manual/rl-next/linux-sandbox-launch-overhead.md

@@ -0,0 +1,12 @@
+---
+synopsis: "Linux sandbox launch overhead greatly reduced"
+cls: [5030]
+category: "Improvements"
+credits: [horrors]
+---
+
+Sandboxed builds are now much cheaper to launch on Linux with 50% lower management
+overhead. This will mostly be noticeable when building derivation trees containing
+many small derivations like nixpkgs' `writeFile` or `runCommand` with scripts that
+very quickly. In synthetic tests we have seen build times of 3000 small runCommand
+drop from 80 seconds to 44 seconds, which is the most optimistic case in practice.

lix/libstore/platform/linux.cc

@@ -1471,15 +1471,10 @@ Pid LinuxLocalDerivationGoal::startChild(
         return {std::move(userns), std::move(netns)};
     }();
 
-    Pipe sendPid;
-    sendPid.create();
-
-    Pid helper = startProcess([&]() {
+    Pid pid = inVFork(/* flags*/ 0, [&]() {
         if (prctl(PR_SET_PDEATHSIG, SIGKILL) == -1)
             throw SysError("setting death signal");
 
-        sendPid.readSide.close();
-
         if (dup2(logPTY.get(), STDERR_FILENO) == -1) {
             throw SysError("failed to redirect build output to log file");
         }
@@ -1504,25 +1499,17 @@ Pid LinuxLocalDerivationGoal::startChild(
         options.cloneFlags =
             CLONE_NEWPID | CLONE_NEWNS | CLONE_NEWIPC | CLONE_NEWUTS | CLONE_PARENT | SIGCHLD;
 
-        pid_t child = startProcess([&]() {
-            if (prctl(PR_SET_PDEATHSIG, SIGKILL) == -1)
-                throw SysError("setting death signal");
-            runChild(builder, envStrs, args);
-        }, options).release();
-
-        writeFull(sendPid.writeSide.get(), fmt("%d\n", child));
-        _exit(0);
+        return startProcess(
+            [&]() {
+                if (prctl(PR_SET_PDEATHSIG, SIGKILL) == -1) {
+                    throw SysError("setting death signal");
+                }
+                runChild(builder, envStrs, args);
+            },
+            options
+        );
     });
 
-    sendPid.writeSide.close();
-
-    if (helper.wait() != 0)
-        throw Error("unable to start build process");
-
-    auto ss = tokenizeString<std::vector<std::string>>(readLine(sendPid.readSide.get()));
-    assert(ss.size() == 1);
-    Pid pid = Pid{string2Int<pid_t>(ss[0]).value()};
-
     /* Migrate the child inside the available control group. */
     if (context.cgroup) {
         context.cgroup->adoptProcess(pid.get());

tests/functional/supplementary-groups.sh

@@ -21,13 +21,13 @@ unshare --mount --map-root-user bash <<EOF
   # Fails with default setting
   # TODO better error
   setLocalStore store1
-  expectStderr 1 "\${cmd[@]}" | grepQuiet "unable to start build process"
+  expectStderr 1 "\${cmd[@]}" | grepQuiet "setgroups failed"
 
   # Fails with `require-drop-supplementary-groups`
   # TODO better error
   setLocalStore store2
   NIX_CONFIG='require-drop-supplementary-groups = true' \
-    expectStderr 1 "\${cmd[@]}" | grepQuiet "unable to start build process"
+    expectStderr 1 "\${cmd[@]}" | grepQuiet "setgroups failed"
 
   # Works without `require-drop-supplementary-groups`
   setLocalStore store3