Piotr Mlocekanddivesh
604b9f66f6
Supervisors poll the gateway for configuration every 10 seconds. This PR adds an opt-in push mode (config_delivery_mode = "push"): supervisors still start by polling, and once their ConnectSupervisor session is accepted, the gateway delivers every later change over it and supervisors apply it and report the result. poll stays the default, and older supervisors and gateways keep polling. --------- Signed-off-by: Piotr Mlocek <pmlocek@nvidia.com> Signed-off-by: divesh <dgude@nvidia.com> Co-authored-by: divesh <dgude@nvidia.com>
OpenShell process supervisor
This crate manages workload processes and SSH sessions through the admitted isolation backend.
Bounded streaming uploads
The CLI's built-in uploads use the openshell-upload-v1 SSH subsystem. Russh
returns SSH receive credit before the supervisor consumes channel data, so this
subsystem returns its own credit and uses SSH for everything else. The gateway
relays the channel without interpreting it.
The client sends one JSON plan line with the destination, then the raw tar
stream, then SSH EOF. When the plan arrives, the supervisor starts the
extractor through the ordinary BoundaryExec path, with the same workload
identity, environment, working directory and filesystem policy as other SSH
children. Startup runs inline, like an exec request.
Stdout carries only big-endian u64 totals of archive bytes that InputTask
has written through BoundaryInput::write_all. The first total, zero, means
extraction started. Extractor stdout and stderr are both sent as SSH stderr,
so extractor output cannot forge credit. The client keeps at most 1 MiB beyond
the last total outstanding; the supervisor rejects input beyond that window.
The existing 4 MiB pending-input guard remains active.
After the extractor exits successfully and stdin has drained, the supervisor sends the final total, then the exit status. The client succeeds only when the status is zero and the final total equals the bytes it sent. A malformed plan, credit violation, extended data or EOF before the plan fails the upload with exit status 74. A rejection wins until the supervisor reports the exit result, even after the extractor exits. Rejection, channel close or disconnect terminates the extractor. Extraction writes directly to the destination, so failures can leave partially extracted files; there is no rollback or atomic replacement.
The CLI requires supervisor support for this subsystem before sending archive data. It does not fall back to the previous unpaced exec-stdin upload path. Ordinary exec stdin and SFTP retain their existing input-budget behavior.