Repository navigation
<
8000
link rel="modulepreload" href="https://github.githubassets.com/assets/1h3-fe71912fb6b520a1.js" crossorigin="anonymous" />
vivarium: QEMU agent channel dies after ~150 large answers #1178
Description
On a QEMU guest booted without EFI (agent on virtconsole, /dev/hvc0), the agent channel stops answering after a burst of large answers, and never answers again.
Repro, on an idle guest at a pause (solid-kubernetes airlab, vms.airlab-offline.driver -- --offline --break hold):
for n in range(300):
rc, out = await cp.execute("head -c 110000 /dev/urandom | base64 -w 0", timeout=30)
Died after 159 answers (~150 KB each, ~23 MB in 42 s): guest stopped answering during: head -c 110000 .... Every later call fails the same way.
Seen on the guest after the failure (from a host-PID pod, because the agent no longer answers):
vivarium-agentis alive, idle inepoll_wait, no child process.- It is the only holder of
/dev/hvc0; the tty is raw (-icanon -isig -ixon -echo -opost). hvc0is not a kernel console (console=ttyS0).
So the request is lost between QEMU and the agent, or the reply after it. Found by a phase that polled kubectl get pods -o json once a second; it lost the guest in every run, in the sandbox and by hand.
Assisted-By: Claude Opus 5.5
Activity