From 62576b0e6b5a234aadcdc10262bcf567d4feb498 Mon Sep 17 00:00:00 2001 From: Barakeinav1 Date: Wed, 3 Jun 2026 10:52:16 +0000 Subject: [PATCH] docs: set KillMode=process on the dstack-vmm systemd unit Without KillMode=process the unit defaults to control-group, so restarting dstack-vmm (e.g. needrestart after an unattended-upgrades run) SIGTERMs every qemu CVM in the cgroup and takes all co-located nodes down at once. KillMode=process stops only the daemon, leaving running CVMs alive across a daemon restart. Co-Authored-By: Claude Opus 4.8 (1M context) --- docs/running-an-mpc-node-in-tdx-external-guide.md | 6 ++++++ 1 file changed, 6 insertions(+) diff --git a/docs/running-an-mpc-node-in-tdx-external-guide.md b/docs/running-an-mpc-node-in-tdx-external-guide.md index f2586454c..9b836f2ee 100644 --- a/docs/running-an-mpc-node-in-tdx-external-guide.md +++ b/docs/running-an-mpc-node-in-tdx-external-guide.md @@ -235,6 +235,12 @@ Restart=on-failure RestartSec=5 User=mpc Group=mpc +# Only stop the dstack-vmm process itself — not the qemu CVMs it supervises. +# Without this, systemd's default KillMode=control-group sends SIGTERM to every +# process in the unit's cgroup (i.e. all CVMs) whenever dstack-vmm is restarted — +# for example when an OS package upgrade triggers `needrestart` — taking every +# node on the host down at once. +KillMode=process [Install] WantedBy=multi-user.target