Exec in place instead of spawning and waiting in the CLI's wrapper processes

Aberta
  • Algo errado?
#3,106 1 comentário 0 reações 1 responsável Ver no GitHub

Mantenedores costumam responder em até 1 dia

@fengmk2 já está trabalhando nisso.

Desde 18/10/2026.

Avaliação

Esta issue ainda não foi avaliada.

Descrição

Follow-up to #2936, which made the project-local vpx bin replace itself with the script's Node.js through process.execve() on Unix instead of spawning it and waiting. The global CLI's shims already exec in place (exec_tool in crates/vp_global_cli/src/shim/exec.rs). Several other paths still spawn a child only to wait for it and pass its exit status on.

Why

  • One PID and real signals. The wrappers ignore SIGINT so that Ctrl+C reaches the child through the process group, and none of them forwards SIGTERM or SIGHUP. A tool that holds the PID of vp and sends it a signal (kill <pid>, a process manager, a test harness) stops the wrapper, or nothing, while the real process keeps running. SIGUSR1 opens a Node.js wrapper's inspector instead of the tool's, and kills the Rust vp wrapper while the tool keeps running.
  • Exact exit status. A child killed by a signal comes back as a plain exit code (exit_code_from_status, exit_status_from, process.exit(n) in bin.ts), so callers cannot tell SIGINT from exit 130.
  • Fewer idle processes. vp dev through the global CLI runs three processes: vp (Rust) waits for node dist/bin.js, which waits for the Vite Node.js process. An idle Node.js wrapper holds about 50 MB for as long as the dev server runs.

Startup time does not change, since each wrapper still starts before it execs.

Candidates

Strongest first. "After the child" is what the parent still does once the child exits, which an exec must move earlier or keep on the spawn path.

# Command Where After the child
1 global vpr → node <vite-plus>/dist/bin.js run … js_executor.rs run_js_entry → vp_command::execute_with_terminal_guard nothing (shim mode skips the upgrade notice)
2 global vpx pkg@version / -p / -c → package manager dlx vpx.rs → vp_pm_cli resolution/runner.rs run_command → vp_command::run_command nothing
3 vp env exec --node <v> [--npm <v>] <command> commands/env/exec.rs:235-252 the upgrade notice; the shim branch of the same command already execs
4 vp sync-versions with the packaged sidecar commands/sync_versions.rs:32-35 nothing
5 vp exec <command> for one package, without --report-summary binding/src/exec/workspace.rs:254 nothing in that case
6 vp dev, preview, build, test, pack, doc, lint (not --init) binding/src/cli/execution.rs:36-57 nothing; the --init and fmt paths filter output and stay as they are
7 package-manager passthrough in the local CLI (vp install, add, dlx, …, and the local vpx pkg@version) binding/src/cli/mod.rs → vp_pm_cli runner the toolchain hint after vp why only
8 global CLI delegation to JS (vp dev, build, test, lint, fmt, check, pack, run, …) cli.rs → commands/delegate.rs → js_executor.rs the cached upgrade notice (main.rs:597-599)
9 global package-manager passthrough (vp install, add, why, …) cli.rs → vp_pm_cli runner the vp why hint and the upgrade notice

Rows 2, 7, and 9 share vp_pm_cli's runner and vp_command::run_command, so an exec variant there covers them together. Rows 5–7 run inside the Node.js process of the local CLI. For row 8, the upgrade notice has to print before the exec, keep the spawn path when a notice is due, or be dropped as the shim branch already does.

Not candidates: npm install -g / uninstall -g (they link bins afterwards), vp fmt and vp check (they filter or capture output), vp run / vp cache (task graph, many children), multi-package and --parallel vp exec, the installer (it owns the temp directory with the binary), and every call that captures output.

Lessons from #2936

  • Blocking stdio. Node.js sets O_NONBLOCK on a pipe once it opens it as stdio, and process.execve() passes that on (nodejs/node#66824), so the new program's writes fail with EAGAIN. bin.ts already restores blocking stdio; anything that execs from Node.js must do so first.
  • Signal dispositions. Node.js ignores SIGPIPE and SIGXFSZ, and an ignored signal stays ignored across exec. std's CommandExt::exec restores SIGPIPE but not SIGXFSZ. vpx-bin.ts restores both by adding and removing a listener.
  • A failed exec. process.execve() aborts Node.js on failure before 26.12 instead of throwing. vpx-bin.ts execs binaries through /bin/sh -c 'exec "$0" "$@"' for that reason. A Rust exec returns an error the caller can still report.
  • IPC. Exec closes an IPC channel, so a process started with one (NODE_CHANNEL_FD, process.send) keeps spawning.
  • Windows has no exec, so every path keeps its spawn there.
  • Rust side. tokio::process::Command::as_std_mut().exec() keeps the configured env, cwd, and pre_exec hooks (such as fix_stdio_streams), so the existing command builders can be reused.

Suggested order

  1. Add a Unix exec variant to vp_command next to run_command and execute_with_terminal_guard, and use it for rows 1–4.
  2. Use it for the local CLI's paths (rows 5–7), skipping exec when the process has an IPC channel and restoring SIGXFSZ.
  3. Decide the upgrade notice, then switch rows 8 and 9.

Separately, vp run does not forward SIGTERM to its tasks, which is out of scope here because it manages many children.

Linguagem predominante
Rust
Estrelas
6.5k
Forks
297
Merge médio
20h 56min
PRs com merge (30d)
234

Preparar o ambiente

Abrir no Codespaces

Inicia o contêiner de desenvolvimento do projeto no navegador, com a sua própria conta do GitHub.

Primeiros passos

  1. Leia a issue inteira e depois o guia de contribuição do projeto.
  2. Comente na issue dizendo que vai assumir — evita que duas pessoas façam o mesmo trabalho.
  3. Faça um fork do repositório e trabalhe em uma branch.
  4. Abra um pull request que referencie o número da issue.

Mais de voidzero-dev/vite-plus

Todas as issues de voidzero-dev/vite-plus

Issues semelhantes

Mais issues de Rust

Receba novas issues na sua caixa de entrada

Um resumo curto de issues do GitHub para quem está começando.