Repository navigation
brev refresh writes <instance>-host with the instance IP but the gateway port (v0.6.337), so host SSH times out #531
Description
Summary
Since v0.6.331+ (observed on v0.6.337), brev refresh writes the <instance>-host SSH alias with the instance's own public IP as Hostname but the gateway's port as Port. The pair is unreachable, so ssh <instance>-host (and brev exec --host) time out on a healthy VM-mode instance, while ssh <instance> reaches the same VM through the gateway.
Possibly the root cause of #436 ("reaches the published host SSH endpoint but times out").
Observed (v0.6.337, macOS, VM-mode instances)
| Instance | <instance> alias |
<instance>-host alias |
|---|---|---|
| AWS g4dn.xlarge | global.prd.ga.run.brev.nvidia.com:23126 → reachable, lands on the VM (hostname brev-…, nvidia-smi, docker) |
<public IP>:23126 → connect timed out; <public IP>:22 also timed out |
| GCP n2d-standard-16 | global.prd.ga.run.brev.nvidia.com:18041 → TCP accepted |
<public IP>:18041 → refused/timed out |
Inside the AWS VM, sshd listens on 0.0.0.0:22 and checkip returns the same public IP, so the alias names the right host with the wrong port (the gateway's).
Cause
pkg/ssh/sshconfigurer.go makeSSHConfigEntryV2:
- v0.6.324: the
-hostentry usesHostName: hostname(the workspace DNS, i.e. the gateway) withPort: workspace.GetHostSSHPort(). - v0.6.337 (and
main): it usesHostName: workspace.GetHostSSHHostname()(the instance IP) with the sameGetHostSSHPort().
For these instances HostSSHPort == SSHPort is the gateway's port, so the new hostname and the old port no longer belong together.
Expected
<instance>-host reaches the host: either the gateway with HostSSHPort (as in v0.6.324), or the instance address with a port its security group opens.
Workaround
Use the <instance> alias for VM-mode instances.
Activity