feat(vmm): support standalone shared netd deployment

Author: kvinwang

docs/libvirt-network-filter.md

@@ -67,8 +67,38 @@ validated by libvirt.
 
 ## Deployment modes
 
-Production packages may socket-activate one shared `netd`, but that is only a
-deployment convenience. Development mode is two ordinary commands:
+Production should run one shared service. `netd` reads only the `[netd]`
+section, so its root-owned configuration can be small and independent of every
+VMM instance:
+
+```toml
+# /etc/dstack/netd.toml
+[netd]
+socket = "/run/dstack/netd.sock"
+allowed_uids = [991, 992]
+libvirt_uri = "qemu:///system"
+```
+
+```ini
+# /etc/systemd/system/dstack-netd.service
+[Unit]
+Description=dstack host networking service
+After=libvirtd.service
+
+[Service]
+ExecStart=/usr/bin/dstack-vmm --config /etc/dstack/netd.toml netd
+Restart=on-failure
+
+[Install]
+WantedBy=multi-user.target
+```
+
+All VMM instance configurations point to the same socket and use distinct
+`cvm.instance_id` values. A dedicated netd uses a different socket. A
+host-wide lock serializes mutations made by shared and dedicated netd
+processes.
+
+Development mode is two ordinary commands:
 
 ```bash
 sudo dstack-vmm --config ./vmm.toml netd \

dstack/vmm/src/app.rs

@@ -499,7 +499,7 @@ impl App {
         }
         self.set_started(id, false)?;
         self.stop_vm_process(id).await?;
-        let networks = self.work_dir(id).runtime_networks();
+        let networks = self.work_dir(id)?.runtime_networks();
         self.remove_filtered_networks(id, &networks).await?;
         Ok(())
     }
@@ -690,7 +690,7 @@ impl App {
             }
         }
 
-        let runtime_networks = self.work_dir(id).runtime_networks();
+        let runtime_networks = self.work_dir(id)?.runtime_networks();
         if let Err(error) = self.remove_filtered_networks(id, &runtime_networks).await {
             warn!(id, %error, "failed to remove filtered networking during VM removal");
         }

dstack/vmm/src/main.rs

@@ -7,7 +7,7 @@ use std::{path::Path, time::Duration};
 use anyhow::{anyhow, Context, Result};
 use app::App;
 use clap::{Args as ClapArgs, Parser, Subcommand};
-use config::Config;
+use config::{Config, NetdConfig};
 use dstack_api_auth::{Authenticator, HttpAuthConfig, HttpAuthFairing};
 use guest_api_service::GuestApiHandler;
 use host_api_service::HostApiHandler;
@@ -194,23 +194,28 @@ async fn main() -> Result<()> {
     }
 
     let figment = config::load_config_figment(args.config.as_deref());
-    let mut config = Config::extract_or_default(&figment)?.abs_path()?;
-    config.cvm.instance_id = netd::instance_id(&config.cvm.instance_id, config.run_path.as_path());
-    if let Some(socket) = args.netd_socket.as_deref() {
-        config.netd.socket = socket.into();
-    }
-
     if let Some(Command::Netd(netd_args)) = &args.command {
+        let mut netd_config: NetdConfig = figment
+            .extract_inner("netd")
+            .context("failed to load [netd] configuration")?;
+        if let Some(socket) = args.netd_socket.as_deref() {
+            netd_config.socket = socket.into();
+        }
         if let Some(socket) = netd_args.socket.as_deref() {
-            config.netd.socket = socket.into();
+            netd_config.socket = socket.into();
         }
-        config
-            .netd
+        netd_config
             .allowed_uids
             .extend(netd_args.allow_uids.iter().copied());
-        config.netd.allowed_uids.sort_unstable();
-        config.netd.allowed_uids.dedup();
-        return netd::serve(config.netd).await;
+        netd_config.allowed_uids.sort_unstable();
+        netd_config.allowed_uids.dedup();
+        return netd::serve(netd_config).await;
+    }
+
+    let mut config = Config::extract_or_default(&figment)?.abs_path()?;
+    config.cvm.instance_id = netd::instance_id(&config.cvm.instance_id, config.run_path.as_path());
+    if let Some(socket) = args.netd_socket.as_deref() {
+        config.netd.socket = socket.into();
     }
 
     // Preserve the existing startup validation. The broader static checks are

dstack/vmm/src/one_shot.rs

@@ -3,10 +3,10 @@
 // SPDX-License-Identifier: Apache-2.0
 
 use crate::app::{
-    make_sys_config, simulator_config_for_manifest, sync_tee_simulator_config, Image, VmConfig,
-    VmWorkDir,
+    make_sys_config, resolved_networks, simulator_config_for_manifest, sync_tee_simulator_config,
+    Image, VmConfig, VmWorkDir,
 };
-use crate::config::Config;
+use crate::config::{Config, NetworkFilterMode, NetworkingMode};
 use crate::main_service;
 use anyhow::{Context, Result};
 use fs_err as fs;
@@ -279,6 +279,17 @@ Compose file content (first 200 chars):
         gateway_enabled: app_compose.gateway_enabled(),
     };
 
+    if !dry_run
+        && config.cvm.network_filter.mode == NetworkFilterMode::Libvirt
+        && resolved_networks(&manifest, &config.cvm)
+            .iter()
+            .any(|network| network.mode == NetworkingMode::Bridge)
+    {
+        anyhow::bail!(
+            "one-shot execution does not manage libvirt-filtered TAP lifecycle; run the VMM server directly or use --dry-run"
+        );
+    }
+
     let process_configs = vm_builder_config
         .config_qemu(&workdir_path, &config.cvm, &gpus)
         .context("Failed to build QEMU configuration")?;