Repository navigation
Description
Summary
We'd like confirmation of the intended ingress design, so that ACI can align its supported contract and documentation. A downstream SDK consumer needs host-loopback access independently of general ingress, and does not need remote LAN access. That case works today. The open question is general ingress: allow. ACI's documentation says ingress must be deny, but the bundle accepts allow, and OpenVMM then binds published listeners on all host interfaces.
Related: #276 (bidirectional host-loopback access, closed 2026-10-13), and #495.
Versions
| Component | Version |
|---|---|
| Rust crates (crates.io) | aci_edge_sandboxes 0.1.1, aci_edge_sandboxes_model 0.1.1, nvx =0.1.0-alpha.6 |
| Runtime bundle | Azure/aci-edge-sandbox v0.1.0-build188642429, asset aci-edge-sandboxes-0.1.0-build188642429.zip, SHA-256 cad4db7c574a73aaa1c5a07e7a740cd2b01d5b83c62c81b2196d63bad9ecc126 (the same bundle as crate 0.1.0) |
| NVX in the bundle | 6b485c82c9fdf4f592e1221efc7d6d41ab9bc66c |
| OpenVMM in the bundle | c32c57d129a36fc8c82308d6659f8d7ab5955474 (OpenVMM log: version="0.3.0+gc32c57d12") |
| Host | Windows 11 x64, Windows Hypervisor Platform |
| Guest image | mcr.microsoft.com/azurelinux/base/python:3 |
The bundle's agent library (aci_edge_agent.dll) is ACI's sandbox agent. It translates the crate's network policy into NVX/OpenVMM options in an ACI layer on top of nvx, so some behavior below comes from that ACI layer rather than from NVX. Each section says which.
Observed
The guest network was 10.0.0.2/24. The guest listener was a Python TCP or UDP server bound to 0.0.0.0 or 10.0.0.2. Probes ran on the same Windows host.
ingress.default |
ingress.hostLoopback |
OpenVMM arguments | OpenVMM host bind (netstat) |
Host 127.0.0.1:<port> |
Host's own LAN IP :<port> |
|---|---|---|---|---|---|
deny |
allow |
--network-ingress deny --host-loopback allow |
127.0.0.1:<port> |
reachable | not reachable |
allow |
allow |
--network-ingress allow --host-loopback allow |
0.0.0.0:<port> (TCP and UDP) |
reachable | reachable (guest saw the host's LAN address as the peer) |
allow |
unset or deny |
--network-ingress allow --host-loopback deny |
0.0.0.0:<port> |
TCP connects, then hangs or resets | — |
deny |
unset or deny |
— | — | refused | refused |
Same results for cold boot and for a platform-snapshot clone whose network policy was replaced. In every case:
- A listener bound only to guest
127.0.0.1was never published. - Host IPv6 loopback
[::1]:<port>was never reachable.
Not tested: access from a second machine on the LAN. That depends on Windows Firewall: a prompt appears when openvmm.exe binds 0.0.0.0.
From source (c32c57d1, not verified at runtime): vm/devices/net/net_consomme/src/ingress.rs binds the unspecified address for private-network ingress and loopback only for host-loopback publication, and consomme/src/inbound.rs admits only private-network peers.
ACI side: ACI's agent reports ingress_allow = true and passes --network-ingress allow. ACI will align its contract and documentation once the intended semantics are confirmed.
Reproducer
Use the reproducer from the ICMP issue (#1250) with these policies, and a guest command that serves on a port, for example python3 -m http.server 8000 --bind 0.0.0.0. While it runs, check from the host with netstat -ano | findstr :8000 and curl http://127.0.0.1:8000/.
// loopback-only.json: the downstream requirement; works today
{ "egress": { "default": "allow" }, "ingress": { "default": "deny", "hostLoopback": "allow" } }
// general-ingress.json: binds 0.0.0.0 on the host
{ "egress": { "default": "allow" }, "ingress": { "default": "allow", "hostLoopback": "allow" } }Note: the second policy makes openvmm.exe listen on all interfaces, which can raise a Windows Firewall prompt.
Questions
- Can you confirm the intended design: general ingress
allowpublishes listeners to the host's private networks by binding0.0.0.0(admitting only private-network peers), while host-loopback access works independently with general ingress denied, as in Support bidirectional host-loopback access #276? - Should IPv6 host loopback
[::1]be published withhostLoopback: allow, or is IPv4 loopback the intended scope? - Is
ingress: allowtogether with host loopback denied (connects, then hangs) an intended combination, or should it be rejected?
Activity
openvmm: pin the prompt reset of refused ingress connections
doc: specify the ingress and host-loopback contract