chore: release (#87)

Signed-off-by: edera-cultivation[bot] <165992271+edera-cultivation[bot]@users.noreply.github.com>
Co-authored-by: edera-cultivation[bot] <165992271+edera-cultivation[bot]@users.noreply.github.com>

.github/workflows/check.yml

@@ -1,11 +1,19 @@
 name: check
-on: [push, pull_request]
+on:
+  pull_request:
+    branches:
+    - main
+  merge_group:
+    branches:
+    - main
 jobs:
   fmt:
     name: fmt
     runs-on: ubuntu-latest
     steps:
     - uses: actions/checkout@v4
+      with:
+        submodules: recursive
     - uses: dtolnay/rust-toolchain@stable
       with:
         components: rustfmt
@@ -16,4 +24,6 @@ jobs:
     runs-on: ubuntu-latest
     steps:
     - uses: actions/checkout@v4
+      with:
+        submodules: recursive
     - run: ./hack/code/shellcheck.sh

.github/workflows/client.yml

@@ -1,5 +1,11 @@
 name: client
-on: [push, pull_request]
+on:
+  pull_request:
+    branches:
+    - main
+  merge_group:
+    branches:
+    - main
 jobs:
   build:
     strategy:
@@ -16,12 +22,16 @@ jobs:
       TARGET_OS: "${{ matrix.platform.os }}"
       TARGET_ARCH: "${{ matrix.platform.arch }}"
     runs-on: "${{ matrix.platform.on }}"
-    name: build ${{ matrix.platform.os }}-${{ matrix.platform.arch }}
+    name: client build ${{ matrix.platform.os }}-${{ matrix.platform.arch }}
     defaults:
       run:
         shell: bash
     steps:
+    - run: git config --global core.autocrlf false && git config --global core.eol lf
+      if: ${{ matrix.platform.os == 'windows' }}
     - uses: actions/checkout@v4
+      with:
+        submodules: recursive
     - uses: dtolnay/rust-toolchain@stable
       if: ${{ matrix.platform.os != 'darwin' }}
     - uses: dtolnay/rust-toolchain@stable

.github/workflows/kernel.yml

@@ -1,26 +1,30 @@
 name: kernel
 on:
-  push:
-    paths:
-    - "kernel/**"
-    - "hack/ci/**"
   pull_request:
+    branches:
+    - main
     paths:
     - "kernel/**"
     - "hack/ci/**"
+  merge_group:
+    branches:
+    - main
 jobs:
   build:
     runs-on: ubuntu-latest
     strategy:
+      fail-fast: false
       matrix:
         arch:
         - x86_64
         - aarch64
     env:
       TARGET_ARCH: "${{ matrix.arch }}"
-    name: build ${{ matrix.arch }}
+    name: kernel build ${{ matrix.arch }}
     steps:
     - uses: actions/checkout@v4
+      with:
+        submodules: recursive
     - uses: dtolnay/rust-toolchain@stable
     - run: ./hack/ci/install-linux-deps.sh
     - run: ./hack/kernel/build.sh

.github/workflows/nightly.yml

@@ -14,9 +14,11 @@ jobs:
         - aarch64
     env:
       TARGET_ARCH: "${{ matrix.arch }}"
-    name: server ${{ matrix.arch }}
+    name: nightly server ${{ matrix.arch }}
     steps:
     - uses: actions/checkout@v4
+      with:
+        submodules: recursive
     - uses: dtolnay/rust-toolchain@stable
       with:
         targets: "${{ matrix.arch }}-unknown-linux-gnu,${{ matrix.arch }}-unknown-linux-musl"
@@ -68,12 +70,16 @@ jobs:
       TARGET_OS: "${{ matrix.platform.os }}"
       TARGET_ARCH: "${{ matrix.platform.arch }}"
     runs-on: "${{ matrix.platform.on }}"
-    name: client ${{ matrix.platform.os }}-${{ matrix.platform.arch }}
+    name: nightly client ${{ matrix.platform.os }}-${{ matrix.platform.arch }}
     defaults:
       run:
         shell: bash
     steps:
+    - run: git config --global core.autocrlf false && git config --global core.eol lf
+      if: ${{ matrix.platform.os == 'windows' }}
     - uses: actions/checkout@v4
+      with:
+        submodules: recursive
     - uses: dtolnay/rust-toolchain@stable
       if: ${{ matrix.platform.os != 'darwin' }}
     - uses: dtolnay/rust-toolchain@stable
@@ -87,4 +93,10 @@ jobs:
     - uses: actions/upload-artifact@v4
       with:
         name: kratactl-${{ matrix.platform.os }}-${{ matrix.platform.arch }}
-        path: "target/*/release/kratactl*"
+        path: "target/*/release/kratactl"
+      if: ${{ matrix.platform.os != 'windows' }}
+    - uses: actions/upload-artifact@v4
+      with:
+        name: kratactl-${{ matrix.platform.os }}-${{ matrix.platform.arch }}
+        path: "target/*/release/kratactl.exe"
+      if: ${{ matrix.platform.os == 'windows' }}

.github/workflows/os.yml

@@ -1,28 +1,31 @@
 name: os
 on:
-  push:
-    paths:
-    - "os/**"
-    - "hack/os/**"
-    - "hack/ci/**"
   pull_request:
+    branches:
+    - main
     paths:
     - "os/**"
     - "hack/os/**"
     - "hack/ci/**"
+  merge_group:
+    branches:
+    - main
 jobs:
   build:
     runs-on: ubuntu-latest
     strategy:
+      fail-fast: false
       matrix:
         arch:
         - x86_64
         - aarch64
     env:
       TARGET_ARCH: "${{ matrix.arch }}"
-    name: build ${{ matrix.arch }}
+    name: os build ${{ matrix.arch }}
     steps:
     - uses: actions/checkout@v4
+      with:
+        submodules: recursive
     - uses: dtolnay/rust-toolchain@stable
       with:
         targets: "${{ matrix.arch }}-unknown-linux-gnu,${{ matrix.arch }}-unknown-linux-musl"

.github/workflows/release-binaries.yml

@@ -0,0 +1,94 @@
+name: release-binaries
+permissions:
+  contents: write
+on:
+  release:
+    types:
+    - published
+env:
+  CARGO_INCREMENTAL: 0
+  CARGO_NET_GIT_FETCH_WITH_CLI: true
+  CARGO_NET_RETRY: 10
+  CARGO_TERM_COLOR: always
+  RUST_BACKTRACE: 1
+  RUSTUP_MAX_RETRIES: 10
+jobs:
+  server:
+    runs-on: ubuntu-latest
+    strategy:
+      fail-fast: false
+      matrix:
+        arch:
+        - x86_64
+        - aarch64
+    env:
+      TARGET_ARCH: "${{ matrix.arch }}"
+    name: release-binaries server ${{ matrix.arch }}
+    steps:
+    - uses: actions/checkout@v4
+      with:
+        submodules: recursive
+    - uses: dtolnay/rust-toolchain@stable
+      with:
+        targets: "${{ matrix.arch }}-unknown-linux-gnu,${{ matrix.arch }}-unknown-linux-musl"
+    - run: ./hack/ci/install-linux-deps.sh
+    - run: ./hack/dist/bundle.sh
+      env:
+        KRATA_KERNEL_BUILD_JOBS: "5"
+    - run: "./hack/ci/assemble-release-assets.sh bundle-systemd ${{ github.event.release.tag_name }} ${{ matrix.arch }} target/dist/bundle-systemd-${{ matrix.arch }}.tgz"
+    - run: ./hack/dist/deb.sh
+      env:
+        KRATA_KERNEL_BUILD_SKIP: "1"
+    - run: "./hack/ci/assemble-release-assets.sh debian ${{ github.event.release.tag_name }} ${{ matrix.arch }} target/dist/*.deb"
+    - run: ./hack/dist/apk.sh
+      env:
+        KRATA_KERNEL_BUILD_SKIP: "1"
+    - run: "./hack/ci/assemble-release-assets.sh alpine ${{ github.event.release.tag_name }} ${{ matrix.arch }} target/dist/*_${{ matrix.arch }}.apk"
+    - run: ./hack/os/build.sh
+      env:
+        KRATA_KERNEL_BUILD_SKIP: "1"
+    - run: "./hack/ci/assemble-release-assets.sh os ${{ github.event.release.tag_name }} ${{ matrix.arch }} target/os/krata-${{ matrix.arch }}.qcow2"
+    - run: "./hack/ci/upload-release-assets.sh ${{ github.event.release.tag_name }}"
+      env:
+        GITHUB_TOKEN: "${{ secrets.GITHUB_TOKEN }}"
+  client:
+    strategy:
+      fail-fast: false
+      matrix:
+        platform:
+        - { os: linux, arch: x86_64, on: ubuntu-latest, deps: linux }
+        - { os: linux, arch: aarch64, on: ubuntu-latest, deps: linux }
+        - { os: darwin, arch: x86_64, on: macos-14, deps: darwin }
+        - { os: darwin, arch: aarch64, on: macos-14, deps: darwin }
+        - { os: freebsd, arch: x86_64, on: ubuntu-latest, deps: linux }
+        - { os: windows, arch: x86_64, on: windows-latest, deps: windows }
+    env:
+      TARGET_OS: "${{ matrix.platform.os }}"
+      TARGET_ARCH: "${{ matrix.platform.arch }}"
+    runs-on: "${{ matrix.platform.on }}"
+    name: release-binaries client ${{ matrix.platform.os }}-${{ matrix.platform.arch }}
+    defaults:
+      run:
+        shell: bash
+    timeout-minutes: 60
+    steps:
+    - uses: actions/checkout@v4
+      with:
+        submodules: recursive
+    - uses: dtolnay/rust-toolchain@stable
+      if: ${{ matrix.platform.os != 'darwin' }}
+    - uses: dtolnay/rust-toolchain@stable
+      with:
+        targets: "${{ matrix.platform.arch }}-apple-darwin"
+      if: ${{ matrix.platform.os == 'darwin' }}
+    - uses: homebrew/actions/setup-homebrew@master
+      if: ${{ matrix.platform.os == 'darwin' }}
+    - run: ./hack/ci/install-${{ matrix.platform.deps }}-deps.sh
+    - run: ./hack/build/cargo.sh build --release --bin kratactl
+    - run: "./hack/ci/assemble-release-assets.sh kratactl ${{ github.event.release.tag_name }} ${{ matrix.platform.os }}-${{ matrix.platform.arch }} target/*/release/kratactl"
+      if: ${{ matrix.platform.os != 'windows' }}
+    - run: "./hack/ci/assemble-release-assets.sh kratactl ${{ github.event.release.tag_name }} ${{ matrix.platform.os }}-${{ matrix.platform.arch }} target/*/release/kratactl.exe"
+      if: ${{ matrix.platform.os == 'windows' }}
+    - run: "./hack/ci/upload-release-assets.sh ${{ github.event.release.tag_name }}"
+      env:
+        GITHUB_TOKEN: "${{ secrets.GITHUB_TOKEN }}"

.github/workflows/release-plz.yml

@@ -0,0 +1,33 @@
+name: release-plz
+permissions:
+  pull-requests: write
+  contents: write
+on:
+  push:
+    branches:
+    - main
+concurrency:
+  group: "${{ github.workflow }}"
+  cancel-in-progress: true
+jobs:
+  release-plz:
+    name: release-plz
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/create-github-app-token@v1
+        id: generate-token
+        with:
+          app-id: "${{ secrets.EDERA_CULTIVATION_APP_ID }}"
+          private-key: "${{ secrets.EDERA_CULTIVATION_APP_PRIVATE_KEY }}"
+      - uses: actions/checkout@v4
+        with:
+          submodules: recursive
+          fetch-depth: 0
+          token: "${{ steps.generate-token.outputs.token }}"
+      - uses: dtolnay/rust-toolchain@stable
+      - run: ./hack/ci/install-linux-deps.sh
+      - name: release-plz
+        uses: MarcoIeni/release-plz-action@v0.5
+        env:
+          GITHUB_TOKEN: "${{ steps.generate-token.outputs.token }}"
+          CARGO_REGISTRY_TOKEN: "${{ secrets.KRATA_RELEASE_CARGO_TOKEN }}"

.github/workflows/server.yml

@@ -1,18 +1,27 @@
 name: server
-on: [push, pull_request]
+on:
+  pull_request:
+    branches:
+    - main
+  merge_group:
+    branches:
+    - main
 jobs:
   build:
     runs-on: ubuntu-latest
     strategy:
+      fail-fast: false
       matrix:
         arch:
         - x86_64
         - aarch64
     env:
       TARGET_ARCH: "${{ matrix.arch }}"
-    name: build ${{ matrix.arch }}
+    name: server build ${{ matrix.arch }}
     steps:
     - uses: actions/checkout@v4
+      with:
+        submodules: recursive
     - uses: dtolnay/rust-toolchain@stable
     - run: ./hack/ci/install-linux-deps.sh
     - run: ./hack/build/cargo.sh build
@@ -25,9 +34,11 @@ jobs:
         - aarch64
     env:
       TARGET_ARCH: "${{ matrix.arch }}"
-    name: test ${{ matrix.arch }}
+    name: server test ${{ matrix.arch }}
     steps:
     - uses: actions/checkout@v4
+      with:
+        submodules: recursive
     - uses: dtolnay/rust-toolchain@stable
     - run: ./hack/ci/install-linux-deps.sh
     - run: ./hack/build/cargo.sh test
@@ -40,9 +51,11 @@ jobs:
         - aarch64
     env:
       TARGET_ARCH: "${{ matrix.arch }}"
-    name: clippy ${{ matrix.arch }}
+    name: server clippy ${{ matrix.arch }}
     steps:
     - uses: actions/checkout@v4
+      with:
+        submodules: recursive
     - uses: dtolnay/rust-toolchain@stable
       with:
         components: clippy
@@ -57,9 +70,11 @@ jobs:
         - aarch64
     env:
       TARGET_ARCH: "${{ matrix.arch }}"
-    name: initrd ${{ matrix.arch }}
+    name: server initrd ${{ matrix.arch }}
     steps:
     - uses: actions/checkout@v4
+      with:
+        submodules: recursive
     - uses: dtolnay/rust-toolchain@stable
       with:
         targets: "${{ matrix.arch }}-unknown-linux-gnu,${{ matrix.arch }}-unknown-linux-musl"

.gitignore

@@ -1,2 +1 @@
 /target
-Cargo.lock

.gitmodules

@@ -0,0 +1,3 @@
+[submodule "vendor"]
+	path = vendor
+	url = https://github.com/edera-dev/krata-vendor.git

CHANGELOG.md

@@ -0,0 +1,79 @@
+# Changelog
+All notable changes to this project will be documented in this file.
+
+The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/),
+and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
+
+## [Unreleased]
+
+## [0.0.10](https://github.com/edera-dev/krata/compare/v0.0.9...v0.0.10) - 2024-04-22
+
+### Added
+- implement guest exec ([#107](https://github.com/edera-dev/krata/pull/107))
+- implement kernel / initrd oci image support ([#103](https://github.com/edera-dev/krata/pull/103))
+- idm v2 ([#102](https://github.com/edera-dev/krata/pull/102))
+- oci concurrency improvements ([#95](https://github.com/edera-dev/krata/pull/95))
+- oci tar format, bit-perfect disk storage for config and manifest, concurrent image pulls ([#88](https://github.com/edera-dev/krata/pull/88))
+
+### Fixed
+- oci cache store should fallback to copy when rename won't work ([#96](https://github.com/edera-dev/krata/pull/96))
+
+### Other
+- update Cargo.lock dependencies
+
+## [0.0.9](https://github.com/edera-dev/krata/compare/v0.0.8...v0.0.9) - 2024-04-15
+
+### Added
+- oci compliance work ([#85](https://github.com/edera-dev/krata/pull/85))
+- oci packer can now use mksquashfs if available ([#70](https://github.com/edera-dev/krata/pull/70))
+- basic kratactl top command ([#72](https://github.com/edera-dev/krata/pull/72))
+- idm snooping ([#71](https://github.com/edera-dev/krata/pull/71))
+- implement oci image progress ([#64](https://github.com/edera-dev/krata/pull/64))
+- guest metrics support ([#46](https://github.com/edera-dev/krata/pull/46))
+
+### Other
+- init: default to xterm if TERM is not set ([#52](https://github.com/edera-dev/krata/pull/52))
+- update Cargo.toml dependencies
+
+## [0.0.8](https://github.com/edera-dev/krata/compare/v0.0.7...v0.0.8) - 2024-04-09
+
+### Other
+- update Cargo.lock dependencies
+
+## [0.0.7](https://github.com/edera-dev/krata/compare/v0.0.6...v0.0.7) - 2024-04-09
+
+### Other
+- update Cargo.toml dependencies
+- update Cargo.lock dependencies
+
+## [0.0.6](https://github.com/edera-dev/krata/compare/v0.0.5...v0.0.6) - 2024-04-09
+
+### Fixed
+- increase channel acquisition timeout to support lower performance hosts ([#36](https://github.com/edera-dev/krata/pull/36))
+
+### Other
+- update Cargo.toml dependencies
+- update Cargo.lock dependencies
+
+## [0.0.5](https://github.com/edera-dev/krata/compare/v0.0.4...v0.0.5) - 2024-04-09
+
+### Added
+- *(ctl)* add help and about to commands and arguments ([#25](https://github.com/edera-dev/krata/pull/25))
+
+### Other
+- update Cargo.toml dependencies
+- update Cargo.lock dependencies
+
+## [0.0.4](https://github.com/edera-dev/krata/releases/tag/v${version}) - 2024-04-03
+
+### Other
+- implement automatic releases
+- reimplement console to utilize channels, and provide logs support
+- set hostname from launch config
+- implement event stream retries
+- work on parallel reconciliation
+- implement parallel guest reconciliation
+- log when a guest start failures occurs
+- remove device restriction
+- setup loopback interface
+- place running tasks in cgroup

Cargo.toml

@@ -16,7 +16,7 @@ members = [
 resolver = "2"
 
 [workspace.package]
-version = "0.0.3"
+version = "0.0.10"
 homepage = "https://krata.dev"
 license = "Apache-2.0"
 repository = "https://github.com/edera-dev/krata"
@@ -24,20 +24,27 @@ repository = "https://github.com/edera-dev/krata"
 [workspace.dependencies]
 anyhow = "1.0"
 arrayvec = "0.7.4"
-async-compression = "0.4.6"
+async-compression = "0.4.8"
 async-stream = "0.3.5"
-async-trait = "0.1.77"
+async-trait = "0.1.80"
 backhand = "0.15.0"
+base64 = "0.22.0"
 byteorder = "1"
 bytes = "1.5.0"
-cli-tables = "0.2.1"
+cgroups-rs = "0.3.4"
+circular-buffer = "0.1.7"
+comfy-table = "7.1.1"
 crossterm = "0.27.0"
 ctrlc = "3.4.4"
 elf = "0.7.4"
 env_logger = "0.11.0"
-etherparse = "0.14.2"
+etherparse = "0.14.3"
+fancy-duration = "0.9.2"
 flate2 = "1.0"
 futures = "0.3.30"
+human_bytes = "0.4"
+indexmap = "2.2.6"
+indicatif = "0.17.8"
 ipnetwork = "0.20.0"
 libc = "0.2"
 log = "0.4.20"
@@ -50,20 +57,25 @@ oci-spec = "0.6.4"
 once_cell = "1.19.0"
 path-absolutize = "3.1.1"
 path-clean = "1.0.1"
-prost = "0.12.3"
-prost-build = "0.12.3"
+prost = "0.12.4"
+prost-build = "0.12.4"
 prost-reflect-build = "0.13.0"
+prost-types = "0.12.4"
 rand = "0.8.5"
-redb = "2.0.0"
+ratatui = "0.26.2"
+redb = "2.1.0"
 rtnetlink = "0.14.1"
-serde_json = "1.0.113"
+scopeguard = "1.2.0"
+serde_json = "1.0.116"
 serde_yaml = "0.9"
 sha256 = "1.5.0"
 signal-hook = "0.3.17"
 slice-copy = "0.3.0"
 smoltcp = "0.11.0"
+sysinfo = "0.30.11"
+termtree = "0.4.1"
 thiserror = "1.0"
-tokio-tun = "0.11.2"
+tokio-tun = "0.11.4"
 tonic-build = "0.11.0"
 tower = "0.4.13"
 udp-stream = "0.0.11"
@@ -76,16 +88,16 @@ version = "4.4.18"
 features = ["derive"]
 
 [workspace.dependencies.prost-reflect]
-version = "0.13.0"
+version = "0.13.1"
 features = ["derive"]
 
 [workspace.dependencies.reqwest]
-version = "0.12.0"
+version = "0.12.4"
 default-features = false
 features = ["rustls-tls"]
 
 [workspace.dependencies.serde]
-version = "1.0.196"
+version = "1.0.198"
 features = ["derive"]
 
 [workspace.dependencies.sys-mount]
@@ -107,3 +119,7 @@ features = ["tls"]
 [workspace.dependencies.uuid]
 version = "1.6.1"
 features = ["v4"]
+
+[profile.release]
+lto = "fat"
+strip = "symbols"

DEV.md

@@ -28,10 +28,21 @@ it's corresponding code path from the above table.
 
 1. Install the specified Debian version on a x86_64 host _capable_ of KVM (NOTE: KVM is not used, Xen is a type-1 hypervisor).
 
-2. Install required packages: `apt install git xen-system-amd64 flex bison libelf-dev libssl-dev bc`
+2. Install required packages:
+
+```sh
+$ apt install git xen-system-amd64 build-essential libclang-dev musl-tools flex bison libelf-dev libssl-dev bc protobuf-compiler libprotobuf-dev squashfs-tools erofs-utils
+```
 
 3. Install [rustup](https://rustup.rs) for managing a Rust environment.
 
+Make sure to install the targets that you need for krata:
+
+```sh
+$ rustup target add x86_64-unknown-linux-gnu
+$ rustup target add x86_64-unknown-linux-musl
+```
+
 4. Configure `/etc/default/grub.d/xen.cfg` to give krata guests some room:
 
 ```sh
@@ -43,7 +54,7 @@ After changing the grub config, update grub: `update-grub`
 
 Then reboot to boot the system as a Xen dom0.
 
-You can validate that Xen is setup by running `xl info` and ensuring it returns useful information about the Xen hypervisor.
+You can validate that Xen is setup by running `dmesg | grep "Hypervisor detected"` and ensuring it returns a line like `Hypervisor detected: Xen PV`, if that is missing, the host is not running under Xen.
 
 5. Clone the krata source code:
 ```sh
@@ -58,8 +69,8 @@ $ ./hack/kernel/build.sh
 ```
 
 7. Copy the guest kernel image at `target/kernel/kernel-x86_64` to `/var/lib/krata/guest/kernel` to have it automatically detected by kratad.
-8. Launch `./hack/debug/kratanet.sh` and keep it running in the foreground.
-9. Launch `./hack/debug/kratad.sh` and keep it running in the foreground.
+8. Launch `./hack/debug/kratad.sh` and keep it running in the foreground.
+9. Launch `./hack/debug/kratanet.sh` and keep it running in the foreground.
 10. Run kratactl to launch a guest:
 
 ```sh

FAQ.md

@@ -2,7 +2,7 @@
 
 ## How does krata currently work?
 
-The krata hypervisor makes it possible to launch OCI containers on a Xen hypervisor without utilizing the Xen userspace tooling. krata contains just enough of the userspace of Xen (reimplemented in Rust) to start an x86_64 Xen Linux PV guest, and implements a Linux init process that can boot an OCI container. It does so by converting an OCI image into a squashfs file and packaging basic startup data in a bundle which the init container can read.
+The krata hypervisor makes it possible to launch OCI containers on a Xen hypervisor without utilizing the Xen userspace tooling. krata contains just enough of the userspace of Xen (reimplemented in Rust) to start an x86_64 Xen Linux PV guest, and implements a Linux init process that can boot an OCI container. It does so by converting an OCI image into a squashfs/erofs file and packaging basic startup data in a bundle which the init container can read.
 
 In addition, due to the desire to reduce dependence on the dom0 network, krata contains a networking daemon called kratanet. kratanet listens for krata guests to startup and launches a userspace networking environment. krata guests can access the dom0 networking stack via the proxynat layer that makes it possible to communicate over UDP, TCP, and ICMP (echo only) to the outside world. In addition, each krata guest is provided a "gateway" IP (both in IPv4 and IPv6) which utilizes smoltcp to provide a virtual host. That virtual host in the future could dial connections into the container to access container networking resources.
 

README.md

@@ -16,4 +16,13 @@ The Edera Hypervisor
 
 ## Introduction
 
-krata is a single-host hypervisor service built primarily for OCI containers.
+krata is a single-host hypervisor service built for OCI-compliant containers. It isolates containers using a type-1 hypervisor, providing workload isolation that can exceed the security level of KVM-based OCI-compliant runtimes.
+
+krata utilizes the core of the Xen hypervisor, with a fully memory-safe Rust control plane to bring Xen tooling into a new secure era.
+
+## Hardware Support
+
+| Architecture | Completion Level | Virtualization Technology |
+| ------------ | ---------------- | ------------------------- |
+| x86_64       | 100% Completed   | Intel VT-x, AMD-V         |
+| aarch64      | 30% Completed    | AArch64 virtualization    |

crates/ctl/Cargo.toml

@@ -2,7 +2,7 @@
 name = "krata-ctl"
 description = "Command-line tool to control the krata hypervisor"
 license.workspace = true
-version= "0.0.3"
+version.workspace = true
 homepage.workspace = true
 repository.workspace = true
 edition = "2021"
@@ -11,16 +11,24 @@ resolver = "2"
 [dependencies]
 anyhow = { workspace = true }
 async-stream = { workspace = true }
+base64 = { workspace = true }
 clap = { workspace = true }
-cli-tables = { workspace = true }
-crossterm = { workspace = true }
+comfy-table = { workspace = true }
+crossterm = { workspace = true, features = ["event-stream"] }
 ctrlc = { workspace = true, features = ["termination"] }
 env_logger = { workspace = true }
-krata = { path = "../krata", version = "^0.0.3" }
+fancy-duration = { workspace = true }
+human_bytes = { workspace = true }
+indicatif = { workspace = true }
+krata = { path = "../krata", version = "^0.0.10" }
 log = { workspace = true }
 prost-reflect = { workspace = true, features = ["serde"] }
+prost-types = { workspace = true }
+ratatui = { workspace = true }
+serde = { workspace = true }
 serde_json = { workspace = true }
 serde_yaml = { workspace = true }
+termtree = { workspace = true }
 tokio = { workspace = true }
 tokio-stream = { workspace = true }
 tonic = { workspace = true }

crates/ctl/src/cli/attach.rs

@@ -10,8 +10,9 @@ use crate::console::StdioConsoleStream;
 use super::resolve_guest;
 
 #[derive(Parser)]
+#[command(about = "Attach to the guest console")]
 pub struct AttachCommand {
-    #[arg()]
+    #[arg(help = "Guest to attach to, either the name or the uuid")]
     guest: String,
 }
 

crates/ctl/src/cli/destroy.rs

@@ -17,10 +17,15 @@ use tonic::{transport::Channel, Request};
 use crate::cli::resolve_guest;
 
 #[derive(Parser)]
+#[command(about = "Destroy a guest")]
 pub struct DestroyCommand {
-    #[arg(short = 'W', long)]
+    #[arg(
+        short = 'W',
+        long,
+        help = "Wait for the destruction of the guest to complete"
+    )]
     wait: bool,
-    #[arg()]
+    #[arg(help = "Guest to destroy, either the name or the uuid")]
     guest: String,
 }
 
@@ -47,34 +52,31 @@ impl DestroyCommand {
 async fn wait_guest_destroyed(id: &str, events: EventStream) -> Result<()> {
     let mut stream = events.subscribe();
     while let Ok(event) = stream.recv().await {
-        match event {
-            Event::GuestChanged(changed) => {
-                let Some(guest) = changed.guest else {
-                    continue;
-                };
+        let Event::GuestChanged(changed) = event;
+        let Some(guest) = changed.guest else {
+            continue;
+        };
 
-                if guest.id != id {
-                    continue;
-                }
+        if guest.id != id {
+            continue;
+        }
 
-                let Some(state) = guest.state else {
-                    continue;
-                };
+        let Some(state) = guest.state else {
+            continue;
+        };
 
-                if let Some(ref error) = state.error_info {
-                    if state.status() == GuestStatus::Failed {
-                        error!("destroy failed: {}", error.message);
-                        std::process::exit(1);
-                    } else {
-                        error!("guest error: {}", error.message);
-                    }
-                }
-
-                if state.status() == GuestStatus::Destroyed {
-                    std::process::exit(0);
-                }
+        if let Some(ref error) = state.error_info {
+            if state.status() == GuestStatus::Failed {
+                error!("destroy failed: {}", error.message);
+                std::process::exit(1);
+            } else {
+                error!("guest error: {}", error.message);
             }
         }
+
+        if state.status() == GuestStatus::Destroyed {
+            std::process::exit(0);
+        }
     }
     Ok(())
 }

crates/ctl/src/cli/exec.rs

@@ -0,0 +1,70 @@
+use std::collections::HashMap;
+
+use anyhow::Result;
+
+use clap::Parser;
+use krata::v1::{
+    common::{GuestTaskSpec, GuestTaskSpecEnvVar},
+    control::{control_service_client::ControlServiceClient, ExecGuestRequest},
+};
+
+use tonic::{transport::Channel, Request};
+
+use crate::console::StdioConsoleStream;
+
+use super::resolve_guest;
+
+#[derive(Parser)]
+#[command(about = "Execute a command inside the guest")]
+pub struct ExecCommand {
+    #[arg[short, long, help = "Environment variables"]]
+    env: Option<Vec<String>>,
+    #[arg(short = 'w', long, help = "Working directory")]
+    working_directory: Option<String>,
+    #[arg(help = "Guest to exec inside, either the name or the uuid")]
+    guest: String,
+    #[arg(
+        allow_hyphen_values = true,
+        trailing_var_arg = true,
+        help = "Command to run inside the guest"
+    )]
+    command: Vec<String>,
+}
+
+impl ExecCommand {
+    pub async fn run(self, mut client: ControlServiceClient<Channel>) -> Result<()> {
+        let guest_id: String = resolve_guest(&mut client, &self.guest).await?;
+        let initial = ExecGuestRequest {
+            guest_id,
+            task: Some(GuestTaskSpec {
+                environment: env_map(&self.env.unwrap_or_default())
+                    .iter()
+                    .map(|(key, value)| GuestTaskSpecEnvVar {
+                        key: key.clone(),
+                        value: value.clone(),
+                    })
+                    .collect(),
+                command: self.command,
+                working_directory: self.working_directory.unwrap_or_default(),
+            }),
+            data: vec![],
+        };
+
+        let stream = StdioConsoleStream::stdin_stream_exec(initial).await;
+
+        let response = client.exec_guest(Request::new(stream)).await?.into_inner();
+
+        let code = StdioConsoleStream::exec_output(response).await?;
+        std::process::exit(code);
+    }
+}
+
+fn env_map(env: &[String]) -> HashMap<String, String> {
+    let mut map = HashMap::<String, String>::new();
+    for item in env {
+        if let Some((key, value)) = item.split_once('=') {
+            map.insert(key.to_string(), value.to_string());
+        }
+    }
+    map
+}

crates/ctl/src/cli/identify_host.rs

@@ -0,0 +1,22 @@
+use anyhow::Result;
+use clap::Parser;
+use krata::v1::control::{control_service_client::ControlServiceClient, IdentifyHostRequest};
+
+use tonic::{transport::Channel, Request};
+
+#[derive(Parser)]
+#[command(about = "Identify information about the host")]
+pub struct IdentifyHostCommand {}
+
+impl IdentifyHostCommand {
+    pub async fn run(self, mut client: ControlServiceClient<Channel>) -> Result<()> {
+        let response = client
+            .identify_host(Request::new(IdentifyHostRequest {}))
+            .await?
+            .into_inner();
+        println!("Host UUID: {}", response.host_uuid);
+        println!("Host Domain: {}", response.host_domid);
+        println!("Krata Version: {}", response.krata_version);
+        Ok(())
+    }
+}

crates/ctl/src/cli/idm_snoop.rs

@@ -0,0 +1,157 @@
+use anyhow::Result;
+use base64::Engine;
+use clap::{Parser, ValueEnum};
+use krata::{
+    events::EventStream,
+    idm::{internal, serialize::IdmSerializable, transport::IdmTransportPacketForm},
+    v1::control::{control_service_client::ControlServiceClient, SnoopIdmReply, SnoopIdmRequest},
+};
+
+use serde::{Deserialize, Serialize};
+use serde_json::Value;
+use tokio_stream::StreamExt;
+use tonic::transport::Channel;
+
+use crate::format::{kv2line, proto2dynamic, value2kv};
+
+#[derive(ValueEnum, Clone, Debug, PartialEq, Eq)]
+enum IdmSnoopFormat {
+    Simple,
+    Jsonl,
+    KeyValue,
+}
+
+#[derive(Parser)]
+#[command(about = "Snoop on the IDM bus")]
+pub struct IdmSnoopCommand {
+    #[arg(short, long, default_value = "simple", help = "Output format")]
+    format: IdmSnoopFormat,
+}
+
+impl IdmSnoopCommand {
+    pub async fn run(
+        self,
+        mut client: ControlServiceClient<Channel>,
+        _events: EventStream,
+    ) -> Result<()> {
+        let mut stream = client.snoop_idm(SnoopIdmRequest {}).await?.into_inner();
+
+        while let Some(reply) = stream.next().await {
+            let reply = reply?;
+            let Some(line) = convert_idm_snoop(reply) else {
+                continue;
+            };
+
+            match self.format {
+                IdmSnoopFormat::Simple => {
+                    self.print_simple(line)?;
+                }
+
+                IdmSnoopFormat::Jsonl => {
+                    let encoded = serde_json::to_string(&line)?;
+                    println!("{}", encoded.trim());
+                }
+
+                IdmSnoopFormat::KeyValue => {
+                    self.print_key_value(line)?;
+                }
+            }
+        }
+
+        Ok(())
+    }
+
+    fn print_simple(&self, line: IdmSnoopLine) -> Result<()> {
+        let encoded = if !line.packet.decoded.is_null() {
+            serde_json::to_string(&line.packet.decoded)?
+        } else {
+            base64::prelude::BASE64_STANDARD.encode(&line.packet.data)
+        };
+        println!(
+            "({} -> {}) {} {} {}",
+            line.from, line.to, line.packet.id, line.packet.form, encoded
+        );
+        Ok(())
+    }
+
+    fn print_key_value(&self, line: IdmSnoopLine) -> Result<()> {
+        let kvs = value2kv(serde_json::to_value(line)?)?;
+        println!("{}", kv2line(kvs));
+        Ok(())
+    }
+}
+
+#[derive(Serialize, Deserialize)]
+pub struct IdmSnoopLine {
+    pub from: String,
+    pub to: String,
+    pub packet: IdmSnoopData,
+}
+
+#[derive(Serialize, Deserialize)]
+pub struct IdmSnoopData {
+    pub id: u64,
+    pub channel: u64,
+    pub form: String,
+    pub data: String,
+    pub decoded: Value,
+}
+
+pub fn convert_idm_snoop(reply: SnoopIdmReply) -> Option<IdmSnoopLine> {
+    let packet = &(reply.packet?);
+
+    let decoded = if packet.channel == 0 {
+        match packet.form() {
+            IdmTransportPacketForm::Event => internal::Event::decode(&packet.data)
+                .ok()
+                .and_then(|event| proto2dynamic(event).ok()),
+
+            IdmTransportPacketForm::Request
+            | IdmTransportPacketForm::StreamRequest
+            | IdmTransportPacketForm::StreamRequestUpdate => {
+                internal::Request::decode(&packet.data)
+                    .ok()
+                    .and_then(|event| proto2dynamic(event).ok())
+            }
+
+            IdmTransportPacketForm::Response | IdmTransportPacketForm::StreamResponseUpdate => {
+                internal::Response::decode(&packet.data)
+                    .ok()
+                    .and_then(|event| proto2dynamic(event).ok())
+            }
+
+            _ => None,
+        }
+    } else {
+        None
+    };
+
+    let decoded = decoded
+        .and_then(|message| serde_json::to_value(message).ok())
+        .unwrap_or(Value::Null);
+
+    let data = IdmSnoopData {
+        id: packet.id,
+        channel: packet.channel,
+        form: match packet.form() {
+            IdmTransportPacketForm::Raw => "raw".to_string(),
+            IdmTransportPacketForm::Event => "event".to_string(),
+            IdmTransportPacketForm::Request => "request".to_string(),
+            IdmTransportPacketForm::Response => "response".to_string(),
+            IdmTransportPacketForm::StreamRequest => "stream-request".to_string(),
+            IdmTransportPacketForm::StreamRequestUpdate => "stream-request-update".to_string(),
+            IdmTransportPacketForm::StreamRequestClosed => "stream-request-closed".to_string(),
+            IdmTransportPacketForm::StreamResponseUpdate => "stream-response-update".to_string(),
+            IdmTransportPacketForm::StreamResponseClosed => "stream-response-closed".to_string(),
+            _ => format!("unknown-{}", packet.form),
+        },
+        data: base64::prelude::BASE64_STANDARD.encode(&packet.data),
+        decoded,
+    };
+
+    Some(IdmSnoopLine {
+        from: reply.from,
+        to: reply.to,
+        packet: data,
+    })
+}

crates/ctl/src/cli/launch.rs

@@ -1,17 +1,17 @@
 use std::collections::HashMap;
 
 use anyhow::Result;
-use clap::Parser;
+use clap::{Parser, ValueEnum};
 use krata::{
     events::EventStream,
     v1::{
         common::{
             guest_image_spec::Image, GuestImageSpec, GuestOciImageSpec, GuestSpec, GuestStatus,
-            GuestTaskSpec, GuestTaskSpecEnvVar,
+            GuestTaskSpec, GuestTaskSpecEnvVar, OciImageFormat,
         },
         control::{
             control_service_client::ControlServiceClient, watch_events_reply::Event,
-            CreateGuestRequest,
+            CreateGuestRequest, PullImageRequest,
         },
     },
 };
@@ -19,40 +19,108 @@ use log::error;
 use tokio::select;
 use tonic::{transport::Channel, Request};
 
-use crate::console::StdioConsoleStream;
+use crate::{console::StdioConsoleStream, pull::pull_interactive_progress};
 
-#[derive(Parser)]
-pub struct LauchCommand {
-    #[arg(short, long)]
-    name: Option<String>,
-    #[arg(short, long, default_value_t = 1)]
-    cpus: u32,
-    #[arg(short, long, default_value_t = 512)]
-    mem: u64,
-    #[arg[short, long]]
-    env: Option<Vec<String>>,
-    #[arg(short, long)]
-    attach: bool,
-    #[arg(short = 'W', long)]
-    wait: bool,
-    #[arg()]
-    oci: String,
-    #[arg(allow_hyphen_values = true, trailing_var_arg = true)]
-    run: Vec<String>,
+#[derive(ValueEnum, Clone, Debug, PartialEq, Eq)]
+pub enum LaunchImageFormat {
+    Squashfs,
+    Erofs,
 }
 
-impl LauchCommand {
+#[derive(Parser)]
+#[command(about = "Launch a new guest")]
+pub struct LaunchCommand {
+    #[arg(long, default_value = "squashfs", help = "Image format")]
+    image_format: LaunchImageFormat,
+    #[arg(long, help = "Overwrite image cache on pull")]
+    pull_overwrite_cache: bool,
+    #[arg(short, long, help = "Name of the guest")]
+    name: Option<String>,
+    #[arg(
+        short,
+        long,
+        default_value_t = 1,
+        help = "vCPUs available to the guest"
+    )]
+    cpus: u32,
+    #[arg(
+        short,
+        long,
+        default_value_t = 512,
+        help = "Memory available to the guest, in megabytes"
+    )]
+    mem: u64,
+    #[arg[short, long, help = "Environment variables set in the guest"]]
+    env: Option<Vec<String>>,
+    #[arg(
+        short,
+        long,
+        help = "Attach to the guest after guest starts, implies --wait"
+    )]
+    attach: bool,
+    #[arg(
+        short = 'W',
+        long,
+        help = "Wait for the guest to start, implied by --attach"
+    )]
+    wait: bool,
+    #[arg(short = 'k', long, help = "OCI kernel image for guest to use")]
+    kernel: Option<String>,
+    #[arg(short = 'I', long, help = "OCI initrd image for guest to use")]
+    initrd: Option<String>,
+    #[arg(short = 'w', long, help = "Working directory")]
+    working_directory: Option<String>,
+    #[arg(help = "Container image for guest to use")]
+    oci: String,
+    #[arg(
+        allow_hyphen_values = true,
+        trailing_var_arg = true,
+        help = "Command to run inside the guest"
+    )]
+    command: Vec<String>,
+}
+
+impl LaunchCommand {
     pub async fn run(
         self,
         mut client: ControlServiceClient<Channel>,
         events: EventStream,
     ) -> Result<()> {
+        let image = self
+            .pull_image(
+                &mut client,
+                &self.oci,
+                match self.image_format {
+                    LaunchImageFormat::Squashfs => OciImageFormat::Squashfs,
+                    LaunchImageFormat::Erofs => OciImageFormat::Erofs,
+                },
+            )
+            .await?;
+
+        let kernel = if let Some(ref kernel) = self.kernel {
+            let kernel_image = self
+                .pull_image(&mut client, kernel, OciImageFormat::Tar)
+                .await?;
+            Some(kernel_image)
+        } else {
+            None
+        };
+
+        let initrd = if let Some(ref initrd) = self.initrd {
+            let kernel_image = self
+                .pull_image(&mut client, initrd, OciImageFormat::Tar)
+                .await?;
+            Some(kernel_image)
+        } else {
+            None
+        };
+
         let request = CreateGuestRequest {
             spec: Some(GuestSpec {
                 name: self.name.unwrap_or_default(),
-                image: Some(GuestImageSpec {
-                    image: Some(Image::Oci(GuestOciImageSpec { image: self.oci })),
-                }),
+                image: Some(image),
+                kernel,
+                initrd,
                 vcpus: self.cpus,
                 mem: self.mem,
                 task: Some(GuestTaskSpec {
@@ -63,7 +131,8 @@ impl LauchCommand {
                             value: value.clone(),
                         })
                         .collect(),
-                    command: self.run,
+                    command: self.command,
+                    working_directory: self.working_directory.unwrap_or_default(),
                 }),
                 annotations: vec![],
             }),
@@ -98,6 +167,28 @@ impl LauchCommand {
         StdioConsoleStream::restore_terminal_mode();
         std::process::exit(code.unwrap_or(0));
     }
+
+    async fn pull_image(
+        &self,
+        client: &mut ControlServiceClient<Channel>,
+        image: &str,
+        format: OciImageFormat,
+    ) -> Result<GuestImageSpec> {
+        let response = client
+            .pull_image(PullImageRequest {
+                image: image.to_string(),
+                format: format.into(),
+                overwrite_cache: self.pull_overwrite_cache,
+            })
+            .await?;
+        let reply = pull_interactive_progress(response.into_inner()).await?;
+        Ok(GuestImageSpec {
+            image: Some(Image::Oci(GuestOciImageSpec {
+                digest: reply.digest,
+                format: reply.format,
+            })),
+        })
+    }
 }
 
 async fn wait_guest_started(id: &str, events: EventStream) -> Result<()> {

crates/ctl/src/cli/list.rs

@@ -1,10 +1,10 @@
 use anyhow::{anyhow, Result};
 use clap::{Parser, ValueEnum};
-use cli_tables::Table;
+use comfy_table::{presets::UTF8_FULL_CONDENSED, Cell, Color, Table};
 use krata::{
     events::EventStream,
     v1::{
-        common::{guest_image_spec::Image, Guest, GuestStatus},
+        common::{Guest, GuestStatus},
         control::{
             control_service_client::ControlServiceClient, ListGuestsRequest, ResolveGuestRequest,
         },
@@ -14,11 +14,11 @@ use krata::{
 use serde_json::Value;
 use tonic::{transport::Channel, Request};
 
-use crate::format::{guest_state_text, guest_status_text, kv2line, proto2dynamic, proto2kv};
+use crate::format::{guest_simple_line, guest_status_text, kv2line, proto2dynamic, proto2kv};
 
 #[derive(ValueEnum, Clone, Debug, PartialEq, Eq)]
 enum ListFormat {
-    CliTable,
+    Table,
     Json,
     JsonPretty,
     Jsonl,
@@ -28,10 +28,11 @@ enum ListFormat {
 }
 
 #[derive(Parser)]
+#[command(about = "List the guests on the hypervisor")]
 pub struct ListCommand {
-    #[arg(short, long, default_value = "cli-table")]
+    #[arg(short, long, default_value = "table", help = "Output format")]
     format: ListFormat,
-    #[arg()]
+    #[arg(help = "Limit to a single guest, either the name or the uuid")]
     guest: Option<String>,
 }
 
@@ -70,24 +71,13 @@ impl ListCommand {
         });
 
         match self.format {
-            ListFormat::CliTable => {
+            ListFormat::Table => {
                 self.print_guest_table(guests)?;
             }
 
             ListFormat::Simple => {
                 for guest in guests {
-                    let state = guest_status_text(
-                        guest
-                            .state
-                            .as_ref()
-                            .map(|x| x.status())
-                            .unwrap_or(GuestStatus::Unknown),
-                    );
-                    let name = guest.spec.as_ref().map(|x| x.name.as_str()).unwrap_or("");
-                    let network = guest.state.as_ref().and_then(|x| x.network.as_ref());
-                    let ipv4 = network.map(|x| x.guest_ipv4.as_str()).unwrap_or("");
-                    let ipv6 = network.map(|x| x.guest_ipv6.as_str()).unwrap_or("");
-                    println!("{}\t{}\t{}\t{}\t{}", guest.id, state, name, ipv4, ipv6);
+                    println!("{}", guest_simple_line(&guest));
                 }
             }
 
@@ -125,49 +115,51 @@ impl ListCommand {
 
     fn print_guest_table(&self, guests: Vec<Guest>) -> Result<()> {
         let mut table = Table::new();
-        let header = vec!["name", "uuid", "state", "ipv4", "ipv6", "image"];
-        table.push_row(&header)?;
+        table.load_preset(UTF8_FULL_CONDENSED);
+        table.set_content_arrangement(comfy_table::ContentArrangement::Dynamic);
+        table.set_header(vec!["name", "uuid", "status", "ipv4", "ipv6"]);
         for guest in guests {
             let ipv4 = guest
                 .state
                 .as_ref()
                 .and_then(|x| x.network.as_ref())
                 .map(|x| x.guest_ipv4.as_str())
-                .unwrap_or("unknown");
+                .unwrap_or("n/a");
             let ipv6 = guest
                 .state
                 .as_ref()
                 .and_then(|x| x.network.as_ref())
                 .map(|x| x.guest_ipv6.as_str())
-                .unwrap_or("unknown");
+                .unwrap_or("n/a");
             let Some(spec) = guest.spec else {
                 continue;
             };
-            let image = spec
-                .image
-                .map(|x| {
-                    x.image
-                        .map(|y| match y {
-                            Image::Oci(oci) => oci.image,
-                        })
-                        .unwrap_or("unknown".to_string())
-                })
-                .unwrap_or("unknown".to_string());
-            table.push_row_string(&vec![
-                spec.name,
-                guest.id,
-                format!("{}", guest_state_text(guest.state.as_ref())),
-                ipv4.to_string(),
-                ipv6.to_string(),
-                image,
-            ])?;
+            let status = guest.state.as_ref().cloned().unwrap_or_default().status();
+            let status_text = guest_status_text(status);
+
+            let status_color = match status {
+                GuestStatus::Destroyed | GuestStatus::Failed => Color::Red,
+                GuestStatus::Destroying | GuestStatus::Exited | GuestStatus::Starting => {
+                    Color::Yellow
+                }
+                GuestStatus::Started => Color::Green,
+                _ => Color::Reset,
+            };
+
+            table.add_row(vec![
+                Cell::new(spec.name),
+                Cell::new(guest.id),
+                Cell::new(status_text).fg(status_color),
+                Cell::new(ipv4.to_string()),
+                Cell::new(ipv6.to_string()),
+            ]);
         }
-        if table.num_records() == 1 {
+        if table.is_empty() {
             if self.guest.is_none() {
                 println!("no guests have been launched");
             }
         } else {
-            println!("{}", table.to_string());
+            println!("{}", table);
         }
         Ok(())
     }

crates/ctl/src/cli/logs.rs

@@ -0,0 +1,58 @@
+use anyhow::Result;
+use async_stream::stream;
+use clap::Parser;
+use krata::{
+    events::EventStream,
+    v1::control::{control_service_client::ControlServiceClient, ConsoleDataRequest},
+};
+
+use tokio::select;
+use tokio_stream::{pending, StreamExt};
+use tonic::transport::Channel;
+
+use crate::console::StdioConsoleStream;
+
+use super::resolve_guest;
+
+#[derive(Parser)]
+#[command(about = "View the logs of a guest")]
+pub struct LogsCommand {
+    #[arg(short, long, help = "Follow output from the guest")]
+    follow: bool,
+    #[arg(help = "Guest to show logs for, either the name or the uuid")]
+    guest: String,
+}
+
+impl LogsCommand {
+    pub async fn run(
+        self,
+        mut client: ControlServiceClient<Channel>,
+        events: EventStream,
+    ) -> Result<()> {
+        let guest_id: String = resolve_guest(&mut client, &self.guest).await?;
+        let guest_id_stream = guest_id.clone();
+        let follow = self.follow;
+        let input = stream! {
+            yield ConsoleDataRequest { guest_id: guest_id_stream, data: Vec::new() };
+            if follow {
+                let mut pending = pending::<ConsoleDataRequest>();
+                while let Some(x) = pending.next().await {
+                    yield x;
+                }
+            }
+        };
+        let output = client.console_data(input).await?.into_inner();
+        let stdout_handle =
+            tokio::task::spawn(async move { StdioConsoleStream::stdout(output).await });
+        let exit_hook_task = StdioConsoleStream::guest_exit_hook(guest_id.clone(), events).await?;
+        let code = select! {
+            x = stdout_handle => {
+                x??;
+                None
+            },
+            x = exit_hook_task => x?
+        };
+        StdioConsoleStream::restore_terminal_mode();
+        std::process::exit(code.unwrap_or(0));
+    }
+}

crates/ctl/src/cli/metrics.rs

@@ -0,0 +1,83 @@
+use anyhow::Result;
+use clap::{Parser, ValueEnum};
+use krata::{
+    events::EventStream,
+    v1::{
+        common::GuestMetricNode,
+        control::{control_service_client::ControlServiceClient, ReadGuestMetricsRequest},
+    },
+};
+
+use tonic::transport::Channel;
+
+use crate::format::{kv2line, metrics_flat, metrics_tree, proto2dynamic};
+
+use super::resolve_guest;
+
+#[derive(ValueEnum, Clone, Debug, PartialEq, Eq)]
+enum MetricsFormat {
+    Tree,
+    Json,
+    JsonPretty,
+    Yaml,
+    KeyValue,
+}
+
+#[derive(Parser)]
+#[command(about = "Read metrics from the guest")]
+pub struct MetricsCommand {
+    #[arg(short, long, default_value = "tree", help = "Output format")]
+    format: MetricsFormat,
+    #[arg(help = "Guest to read metrics for, either the name or the uuid")]
+    guest: String,
+}
+
+impl MetricsCommand {
+    pub async fn run(
+        self,
+        mut client: ControlServiceClient<Channel>,
+        _events: EventStream,
+    ) -> Result<()> {
+        let guest_id: String = resolve_guest(&mut client, &self.guest).await?;
+        let root = client
+            .read_guest_metrics(ReadGuestMetricsRequest { guest_id })
+            .await?
+            .into_inner()
+            .root
+            .unwrap_or_default();
+        match self.format {
+            MetricsFormat::Tree => {
+                self.print_metrics_tree(root)?;
+            }
+
+            MetricsFormat::Json | MetricsFormat::JsonPretty | MetricsFormat::Yaml => {
+                let value = serde_json::to_value(proto2dynamic(root)?)?;
+                let encoded = if self.format == MetricsFormat::JsonPretty {
+                    serde_json::to_string_pretty(&value)?
+                } else if self.format == MetricsFormat::Yaml {
+                    serde_yaml::to_string(&value)?
+                } else {
+                    serde_json::to_string(&value)?
+                };
+                println!("{}", encoded.trim());
+            }
+
+            MetricsFormat::KeyValue => {
+                self.print_key_value(root)?;
+            }
+        }
+
+        Ok(())
+    }
+
+    fn print_metrics_tree(&self, root: GuestMetricNode) -> Result<()> {
+        print!("{}", metrics_tree(root));
+        Ok(())
+    }
+
+    fn print_key_value(&self, metrics: GuestMetricNode) -> Result<()> {
+        let kvs = metrics_flat(metrics);
+        println!("{}", kv2line(kvs));
+        Ok(())
+    }
+}

crates/ctl/src/cli/mod.rs

@@ -1,8 +1,15 @@
 pub mod attach;
 pub mod destroy;
+pub mod exec;
+pub mod identify_host;
+pub mod idm_snoop;
 pub mod launch;
 pub mod list;
+pub mod logs;
+pub mod metrics;
+pub mod pull;
 pub mod resolve;
+pub mod top;
 pub mod watch;
 
 use anyhow::{anyhow, Result};
@@ -10,21 +17,29 @@ use clap::{Parser, Subcommand};
 use krata::{
     client::ControlClientProvider,
     events::EventStream,
-    v1::control::{
-        control_service_client::ControlServiceClient, ResolveGuestRequest, WatchEventsRequest,
-    },
+    v1::control::{control_service_client::ControlServiceClient, ResolveGuestRequest},
 };
 use tonic::{transport::Channel, Request};
 
 use self::{
-    attach::AttachCommand, destroy::DestroyCommand, launch::LauchCommand, list::ListCommand,
-    resolve::ResolveCommand, watch::WatchCommand,
+    attach::AttachCommand, destroy::DestroyCommand, exec::ExecCommand,
+    identify_host::IdentifyHostCommand, idm_snoop::IdmSnoopCommand, launch::LaunchCommand,
+    list::ListCommand, logs::LogsCommand, metrics::MetricsCommand, pull::PullCommand,
+    resolve::ResolveCommand, top::TopCommand, watch::WatchCommand,
 };
 
 #[derive(Parser)]
-#[command(version, about)]
+#[command(
+    version,
+    about = "Control the krata hypervisor, a secure platform for running containers"
+)]
 pub struct ControlCommand {
-    #[arg(short, long, default_value = "unix:///var/lib/krata/daemon.socket")]
+    #[arg(
+        short,
+        long,
+        help = "The connection URL to the krata hypervisor",
+        default_value = "unix:///var/lib/krata/daemon.socket"
+    )]
     connection: String,
 
     #[command(subcommand)]
@@ -33,24 +48,25 @@ pub struct ControlCommand {
 
 #[derive(Subcommand)]
 pub enum Commands {
-    Launch(LauchCommand),
+    Launch(LaunchCommand),
     Destroy(DestroyCommand),
     List(ListCommand),
     Attach(AttachCommand),
+    Pull(PullCommand),
+    Logs(LogsCommand),
     Watch(WatchCommand),
     Resolve(ResolveCommand),
+    Metrics(MetricsCommand),
+    IdmSnoop(IdmSnoopCommand),
+    Top(TopCommand),
+    IdentifyHost(IdentifyHostCommand),
+    Exec(ExecCommand),
 }
 
 impl ControlCommand {
     pub async fn run(self) -> Result<()> {
-        let mut client = ControlClientProvider::dial(self.connection.parse()?).await?;
-        let events = EventStream::open(
-            client
-                .watch_events(WatchEventsRequest {})
-                .await?
-                .into_inner(),
-        )
-        .await?;
+        let client = ControlClientProvider::dial(self.connection.parse()?).await?;
+        let events = EventStream::open(client.clone()).await?;
 
         match self.command {
             Commands::Launch(launch) => {
@@ -65,6 +81,10 @@ impl ControlCommand {
                 attach.run(client, events).await?;
             }
 
+            Commands::Logs(logs) => {
+                logs.run(client, events).await?;
+            }
+
             Commands::List(list) => {
                 list.run(client, events).await?;
             }
@@ -76,6 +96,30 @@ impl ControlCommand {
             Commands::Resolve(resolve) => {
                 resolve.run(client).await?;
             }
+
+            Commands::Metrics(metrics) => {
+                metrics.run(client, events).await?;
+            }
+
+            Commands::IdmSnoop(snoop) => {
+                snoop.run(client, events).await?;
+            }
+
+            Commands::Top(top) => {
+                top.run(client, events).await?;
+            }
+
+            Commands::Pull(pull) => {
+                pull.run(client).await?;
+            }
+
+            Commands::IdentifyHost(identify) => {
+                identify.run(client).await?;
+            }
+
+            Commands::Exec(exec) => {
+                exec.run(client).await?;
+            }
         }
         Ok(())
     }

crates/ctl/src/cli/pull.rs

@@ -0,0 +1,47 @@
+use anyhow::Result;
+use clap::{Parser, ValueEnum};
+use krata::v1::{
+    common::OciImageFormat,
+    control::{control_service_client::ControlServiceClient, PullImageRequest},
+};
+
+use tonic::transport::Channel;
+
+use crate::pull::pull_interactive_progress;
+
+#[derive(ValueEnum, Clone, Debug, PartialEq, Eq)]
+pub enum PullImageFormat {
+    Squashfs,
+    Erofs,
+    Tar,
+}
+
+#[derive(Parser)]
+#[command(about = "Pull an image into the cache")]
+pub struct PullCommand {
+    #[arg(help = "Image name")]
+    image: String,
+    #[arg(short = 's', long, default_value = "squashfs", help = "Image format")]
+    image_format: PullImageFormat,
+    #[arg(short = 'o', long, help = "Overwrite image cache")]
+    overwrite_cache: bool,
+}
+
+impl PullCommand {
+    pub async fn run(self, mut client: ControlServiceClient<Channel>) -> Result<()> {
+        let response = client
+            .pull_image(PullImageRequest {
+                image: self.image.clone(),
+                format: match self.image_format {
+                    PullImageFormat::Squashfs => OciImageFormat::Squashfs.into(),
+                    PullImageFormat::Erofs => OciImageFormat::Erofs.into(),
+                    PullImageFormat::Tar => OciImageFormat::Tar.into(),
+                },
+                overwrite_cache: self.overwrite_cache,
+            })
+            .await?;
+        let reply = pull_interactive_progress(response.into_inner()).await?;
+        println!("{}", reply.digest);
+        Ok(())
+    }
+}

crates/ctl/src/cli/resolve.rs

@@ -5,8 +5,9 @@ use krata::v1::control::{control_service_client::ControlServiceClient, ResolveGu
 use tonic::{transport::Channel, Request};
 
 #[derive(Parser)]
+#[command(about = "Resolve a guest name to a uuid")]
 pub struct ResolveCommand {
-    #[arg()]
+    #[arg(help = "Guest name")]
     guest: String,
 }
 

crates/ctl/src/cli/top.rs

@@ -0,0 +1,215 @@
+use anyhow::Result;
+use clap::Parser;
+use krata::{events::EventStream, v1::control::control_service_client::ControlServiceClient};
+use std::{
+    io::{self, stdout, Stdout},
+    time::Duration,
+};
+use tokio::select;
+use tokio_stream::StreamExt;
+use tonic::transport::Channel;
+
+use crossterm::{
+    event::{Event, KeyCode, KeyEvent, KeyEventKind},
+    execute,
+    terminal::*,
+};
+use ratatui::{
+    prelude::*,
+    symbols::border,
+    widgets::{
+        block::{Position, Title},
+        Block, Borders, Row, Table, TableState,
+    },
+};
+
+use crate::{
+    format::guest_status_text,
+    metrics::{
+        lookup_metric_value, MultiMetricCollector, MultiMetricCollectorHandle, MultiMetricState,
+    },
+};
+
+#[derive(Parser)]
+#[command(about = "Dashboard for running guests")]
+pub struct TopCommand {}
+
+pub type Tui = Terminal<CrosstermBackend<Stdout>>;
+
+impl TopCommand {
+    pub async fn run(
+        self,
+        client: ControlServiceClient<Channel>,
+        events: EventStream,
+    ) -> Result<()> {
+        let collector = MultiMetricCollector::new(client, events, Duration::from_millis(200))?;
+        let collector = collector.launch().await?;
+        let mut tui = TopCommand::init()?;
+        let mut app = TopApp {
+            metrics: MultiMetricState { guests: vec![] },
+            exit: false,
+            table: TableState::new(),
+        };
+        app.run(collector, &mut tui).await?;
+        TopCommand::restore()?;
+        Ok(())
+    }
+
+    pub fn init() -> io::Result<Tui> {
+        execute!(stdout(), EnterAlternateScreen)?;
+        enable_raw_mode()?;
+        Terminal::new(CrosstermBackend::new(stdout()))
+    }
+
+    pub fn restore() -> io::Result<()> {
+        execute!(stdout(), LeaveAlternateScreen)?;
+        disable_raw_mode()?;
+        Ok(())
+    }
+}
+
+pub struct TopApp {
+    table: TableState,
+    metrics: MultiMetricState,
+    exit: bool,
+}
+
+impl TopApp {
+    pub async fn run(
+        &mut self,
+        mut collector: MultiMetricCollectorHandle,
+        terminal: &mut Tui,
+    ) -> Result<()> {
+        let mut events = crossterm::event::EventStream::new();
+
+        while !self.exit {
+            terminal.draw(|frame| self.render_frame(frame))?;
+
+            select! {
+                x = collector.receiver.recv() => match x {
+                    Some(state) => {
+                        self.metrics = state;
+                    },
+
+                    None => {
+                        break;
+                    }
+                },
+
+                x = events.next() => match x {
+                    Some(event) => {
+                        let event = event?;
+                        self.handle_event(event)?;
+                    },
+
+                    None => {
+                        break;
+                    }
+                }
+            };
+        }
+        Ok(())
+    }
+
+    fn render_frame(&mut self, frame: &mut Frame) {
+        frame.render_widget(self, frame.size());
+    }
+
+    fn handle_event(&mut self, event: Event) -> io::Result<()> {
+        match event {
+            Event::Key(key_event) if key_event.kind == KeyEventKind::Press => {
+                self.handle_key_event(key_event)
+            }
+            _ => {}
+        };
+        Ok(())
+    }
+
+    fn exit(&mut self) {
+        self.exit = true;
+    }
+
+    fn handle_key_event(&mut self, key_event: KeyEvent) {
+        if let KeyCode::Char('q') = key_event.code {
+            self.exit()
+        }
+    }
+}
+
+impl Widget for &mut TopApp {
+    fn render(self, area: Rect, buf: &mut Buffer) {
+        let title = Title::from(" krata hypervisor ".bold());
+        let instructions = Title::from(vec![" Quit ".into(), "<Q> ".blue().bold()]);
+        let block = Block::default()
+            .title(title.alignment(Alignment::Center))
+            .title(
+                instructions
+                    .alignment(Alignment::Center)
+                    .position(Position::Bottom),
+            )
+            .borders(Borders::ALL)
+            .border_set(border::THICK);
+
+        let mut rows = vec![];
+
+        for ms in &self.metrics.guests {
+            let Some(ref spec) = ms.guest.spec else {
+                continue;
+            };
+
+            let Some(ref state) = ms.guest.state else {
+                continue;
+            };
+
+            let memory_total = ms
+                .root
+                .as_ref()
+                .and_then(|root| lookup_metric_value(root, "system/memory/total"));
+            let memory_used = ms
+                .root
+                .as_ref()
+                .and_then(|root| lookup_metric_value(root, "system/memory/used"));
+            let memory_free = ms
+                .root
+                .as_ref()
+                .and_then(|root| lookup_metric_value(root, "system/memory/free"));
+
+            let row = Row::new(vec![
+                spec.name.clone(),
+                ms.guest.id.clone(),
+                guest_status_text(state.status()),
+                memory_total.unwrap_or_default(),
+                memory_used.unwrap_or_default(),
+                memory_free.unwrap_or_default(),
+            ]);
+            rows.push(row);
+        }
+
+        let widths = [
+            Constraint::Min(8),
+            Constraint::Min(8),
+            Constraint::Min(8),
+            Constraint::Min(8),
+            Constraint::Min(8),
+            Constraint::Min(8),
+        ];
+
+        let table = Table::new(rows, widths)
+            .header(
+                Row::new(vec![
+                    "name",
+                    "id",
+                    "status",
+                    "total memory",
+                    "used memory",
+                    "free memory",
+                ])
+                .style(Style::new().bold())
+                .bottom_margin(1),
+            )
+            .column_spacing(1)
+            .block(block);
+
+        StatefulWidget::render(table, area, buf, &mut self.table);
+    }
+}

crates/ctl/src/cli/watch.rs

@@ -7,7 +7,7 @@ use krata::{
 use prost_reflect::ReflectMessage;
 use serde_json::Value;
 
-use crate::format::{guest_state_text, kv2line, proto2dynamic, proto2kv};
+use crate::format::{guest_simple_line, kv2line, proto2dynamic, proto2kv};
 
 #[derive(ValueEnum, Clone, Debug, PartialEq, Eq)]
 enum WatchFormat {
@@ -17,8 +17,9 @@ enum WatchFormat {
 }
 
 #[derive(Parser)]
+#[command(about = "Watch for guest changes")]
 pub struct WatchCommand {
-    #[arg(short, long, default_value = "simple")]
+    #[arg(short, long, default_value = "simple", help = "Output format")]
     format: WatchFormat,
 }
 
@@ -27,12 +28,10 @@ impl WatchCommand {
         let mut stream = events.subscribe();
         loop {
             let event = stream.recv().await?;
-            match event {
-                Event::GuestChanged(changed) => {
-                    let guest = changed.guest.clone();
-                    self.print_event("guest.changed", changed, guest)?;
-                }
-            }
+
+            let Event::GuestChanged(changed) = event;
+            let guest = changed.guest.clone();
+            self.print_event("guest.changed", changed, guest)?;
         }
     }
 
@@ -45,12 +44,7 @@ impl WatchCommand {
         match self.format {
             WatchFormat::Simple => {
                 if let Some(guest) = guest {
-                    println!(
-                        "{} guest={} status=\"{}\"",
-                        typ,
-                        guest.id,
-                        guest_state_text(guest.state.as_ref()).replace('"', "\\\"")
-                    );
+                    println!("{}", guest_simple_line(&guest));
                 }
             }
 

crates/ctl/src/console.rs

@@ -1,4 +1,4 @@
-use anyhow::Result;
+use anyhow::{anyhow, Result};
 use async_stream::stream;
 use crossterm::{
     terminal::{disable_raw_mode, enable_raw_mode, is_raw_mode_enabled},
@@ -8,12 +8,15 @@ use krata::{
     events::EventStream,
     v1::{
         common::GuestStatus,
-        control::{watch_events_reply::Event, ConsoleDataReply, ConsoleDataRequest},
+        control::{
+            watch_events_reply::Event, ConsoleDataReply, ConsoleDataRequest, ExecGuestReply,
+            ExecGuestRequest,
+        },
     },
 };
 use log::debug;
 use tokio::{
-    io::{stdin, stdout, AsyncReadExt, AsyncWriteExt},
+    io::{stderr, stdin, stdout, AsyncReadExt, AsyncWriteExt},
     task::JoinHandle,
 };
 use tokio_stream::{Stream, StreamExt};
@@ -45,6 +48,31 @@ impl StdioConsoleStream {
         }
     }
 
+    pub async fn stdin_stream_exec(
+        initial: ExecGuestRequest,
+    ) -> impl Stream<Item = ExecGuestRequest> {
+        let mut stdin = stdin();
+        stream! {
+            yield initial;
+
+            let mut buffer = vec![0u8; 60];
+            loop {
+                let size = match stdin.read(&mut buffer).await {
+                    Ok(size) => size,
+                    Err(error) => {
+                        debug!("failed to read stdin: {}", error);
+                        break;
+                    }
+                };
+                let data = buffer[0..size].to_vec();
+                if size == 1 && buffer[0] == 0x1d {
+                    break;
+                }
+                yield ExecGuestRequest { guest_id: String::default(), task: None, data };
+            }
+        }
+    }
+
     pub async fn stdout(mut stream: Streaming<ConsoleDataReply>) -> Result<()> {
         if stdin().is_tty() {
             enable_raw_mode()?;
@@ -62,6 +90,32 @@ impl StdioConsoleStream {
         Ok(())
     }
 
+    pub async fn exec_output(mut stream: Streaming<ExecGuestReply>) -> Result<i32> {
+        let mut stdout = stdout();
+        let mut stderr = stderr();
+        while let Some(reply) = stream.next().await {
+            let reply = reply?;
+            if !reply.stdout.is_empty() {
+                stdout.write_all(&reply.stdout).await?;
+                stdout.flush().await?;
+            }
+
+            if !reply.stderr.is_empty() {
+                stderr.write_all(&reply.stderr).await?;
+                stderr.flush().await?;
+            }
+
+            if reply.exited {
+                if reply.error.is_empty() {
+                    return Ok(reply.exit_code);
+                } else {
+                    return Err(anyhow!("exec failed: {}", reply.error));
+                }
+            }
+        }
+        Ok(-1)
+    }
+
     pub async fn guest_exit_hook(
         id: String,
         events: EventStream,
@@ -69,29 +123,26 @@ impl StdioConsoleStream {
         Ok(tokio::task::spawn(async move {
             let mut stream = events.subscribe();
             while let Ok(event) = stream.recv().await {
-                match event {
-                    Event::GuestChanged(changed) => {
-                        let Some(guest) = changed.guest else {
-                            continue;
-                        };
+                let Event::GuestChanged(changed) = event;
+                let Some(guest) = changed.guest else {
+                    continue;
+                };
 
-                        let Some(state) = guest.state else {
-                            continue;
-                        };
+                let Some(state) = guest.state else {
+                    continue;
+                };
 
-                        if guest.id != id {
-                            continue;
-                        }
+                if guest.id != id {
+                    continue;
+                }
 
-                        if let Some(exit_info) = state.exit_info {
-                            return Some(exit_info.code);
-                        }
+                if let Some(exit_info) = state.exit_info {
+                    return Some(exit_info.code);
+                }
 
-                        let status = state.status();
-                        if status == GuestStatus::Destroying || status == GuestStatus::Destroyed {
-                            return Some(10);
-                        }
-                    }
+                let status = state.status();
+                if status == GuestStatus::Destroying || status == GuestStatus::Destroyed {
+                    return Some(10);
                 }
             }
             None

crates/ctl/src/format.rs

@@ -1,8 +1,12 @@
-use std::collections::HashMap;
+use std::{collections::HashMap, time::Duration};
 
 use anyhow::Result;
-use krata::v1::common::{GuestState, GuestStatus};
-use prost_reflect::{DynamicMessage, ReflectMessage, Value};
+use fancy_duration::FancyDuration;
+use human_bytes::human_bytes;
+use krata::v1::common::{Guest, GuestMetricFormat, GuestMetricNode, GuestStatus};
+use prost_reflect::{DynamicMessage, ReflectMessage};
+use prost_types::Value;
+use termtree::Tree;
 
 pub fn proto2dynamic(proto: impl ReflectMessage) -> Result<DynamicMessage> {
     Ok(DynamicMessage::decode(
@@ -11,46 +15,59 @@ pub fn proto2dynamic(proto: impl ReflectMessage) -> Result<DynamicMessage> {
     )?)
 }
 
-pub fn proto2kv(proto: impl ReflectMessage) -> Result<HashMap<String, String>> {
-    let message = proto2dynamic(proto)?;
+pub fn value2kv(value: serde_json::Value) -> Result<HashMap<String, String>> {
     let mut map = HashMap::new();
-
-    fn crawl(prefix: &str, map: &mut HashMap<String, String>, message: &DynamicMessage) {
-        for (field, value) in message.fields() {
-            let path = if prefix.is_empty() {
-                field.name().to_string()
+    fn crawl(prefix: String, map: &mut HashMap<String, String>, value: serde_json::Value) {
+        fn dot(prefix: &str, next: String) -> String {
+            if prefix.is_empty() {
+                next.to_string()
             } else {
-                format!("{}.{}", prefix, field.name())
-            };
-            match value {
-                Value::Message(child) => {
-                    crawl(&path, map, child);
-                }
+                format!("{}.{}", prefix, next)
+            }
+        }
 
-                Value::EnumNumber(number) => {
-                    if let Some(e) = field.kind().as_enum() {
-                        if let Some(value) = e.get_value(*number) {
-                            map.insert(path, value.name().to_string());
-                        }
-                    }
-                }
+        match value {
+            serde_json::Value::Null => {
+                map.insert(prefix, "null".to_string());
+            }
 
-                Value::String(value) => {
-                    map.insert(path, value.clone());
-                }
+            serde_json::Value::String(value) => {
+                map.insert(prefix, value);
+            }
 
-                _ => {
-                    map.insert(path, value.to_string());
+            serde_json::Value::Bool(value) => {
+                map.insert(prefix, value.to_string());
+            }
+
+            serde_json::Value::Number(value) => {
+                map.insert(prefix, value.to_string());
+            }
+
+            serde_json::Value::Array(value) => {
+                for (i, item) in value.into_iter().enumerate() {
+                    let next = dot(&prefix, i.to_string());
+                    crawl(next, map, item);
+                }
+            }
+
+            serde_json::Value::Object(value) => {
+                for (key, item) in value {
+                    let next = dot(&prefix, key);
+                    crawl(next, map, item);
                 }
             }
         }
     }
-
-    crawl("", &mut map, &message);
-
+    crawl("".to_string(), &mut map, value);
     Ok(map)
 }
 
+pub fn proto2kv(proto: impl ReflectMessage) -> Result<HashMap<String, String>> {
+    let message = proto2dynamic(proto)?;
+    let value = serde_json::to_value(message)?;
+    value2kv(value)
+}
+
 pub fn kv2line(map: HashMap<String, String>) -> String {
     map.iter()
         .map(|(k, v)| format!("{}=\"{}\"", k, v.replace('"', "\\\"")))
@@ -71,16 +88,77 @@ pub fn guest_status_text(status: GuestStatus) -> String {
     .to_string()
 }
 
-pub fn guest_state_text(state: Option<&GuestState>) -> String {
-    let state = state.cloned().unwrap_or_default();
-    let mut text = guest_status_text(state.status());
-
-    if let Some(exit) = state.exit_info {
-        text.push_str(&format!(" (exit code: {})", exit.code));
-    }
-
-    if let Some(error) = state.error_info {
-        text.push_str(&format!(" (error: {})", error.message));
-    }
-    text
+pub fn guest_simple_line(guest: &Guest) -> String {
+    let state = guest_status_text(
+        guest
+            .state
+            .as_ref()
+            .map(|x| x.status())
+            .unwrap_or(GuestStatus::Unknown),
+    );
+    let name = guest.spec.as_ref().map(|x| x.name.as_str()).unwrap_or("");
+    let network = guest.state.as_ref().and_then(|x| x.network.as_ref());
+    let ipv4 = network.map(|x| x.guest_ipv4.as_str()).unwrap_or("");
+    let ipv6 = network.map(|x| x.guest_ipv6.as_str()).unwrap_or("");
+    format!("{}\t{}\t{}\t{}\t{}", guest.id, state, name, ipv4, ipv6)
+}
+
+fn metrics_value_string(value: Value) -> String {
+    proto2dynamic(value)
+        .map(|x| serde_json::to_string(&x).ok())
+        .ok()
+        .flatten()
+        .unwrap_or_default()
+}
+
+fn metrics_value_numeric(value: Value) -> f64 {
+    let string = metrics_value_string(value);
+    string.parse::<f64>().ok().unwrap_or(f64::NAN)
+}
+
+pub fn metrics_value_pretty(value: Value, format: GuestMetricFormat) -> String {
+    match format {
+        GuestMetricFormat::Bytes => human_bytes(metrics_value_numeric(value)),
+        GuestMetricFormat::Integer => (metrics_value_numeric(value) as u64).to_string(),
+        GuestMetricFormat::DurationSeconds => {
+            FancyDuration(Duration::from_secs_f64(metrics_value_numeric(value))).to_string()
+        }
+        _ => metrics_value_string(value),
+    }
+}
+
+fn metrics_flat_internal(prefix: &str, node: GuestMetricNode, map: &mut HashMap<String, String>) {
+    if let Some(value) = node.value {
+        map.insert(prefix.to_string(), metrics_value_string(value));
+    }
+
+    for child in node.children {
+        let path = if prefix.is_empty() {
+            child.name.to_string()
+        } else {
+            format!("{}.{}", prefix, child.name)
+        };
+        metrics_flat_internal(&path, child, map);
+    }
+}
+
+pub fn metrics_flat(root: GuestMetricNode) -> HashMap<String, String> {
+    let mut map = HashMap::new();
+    metrics_flat_internal("", root, &mut map);
+    map
+}
+
+pub fn metrics_tree(node: GuestMetricNode) -> Tree<String> {
+    let mut name = node.name.to_string();
+    let format = node.format();
+    if let Some(value) = node.value {
+        let value_string = metrics_value_pretty(value, format);
+        name.push_str(&format!(": {}", value_string));
+    }
+
+    let mut tree = Tree::new(name);
+    for child in node.children {
+        tree.push(metrics_tree(child));
+    }
+    tree
 }

crates/ctl/src/lib.rs

@@ -1,3 +1,5 @@
 pub mod cli;
 pub mod console;
 pub mod format;
+pub mod metrics;
+pub mod pull;

crates/ctl/src/metrics.rs

@@ -0,0 +1,158 @@
+use anyhow::Result;
+use krata::{
+    events::EventStream,
+    v1::{
+        common::{Guest, GuestMetricNode, GuestStatus},
+        control::{
+            control_service_client::ControlServiceClient, watch_events_reply::Event,
+            ListGuestsRequest, ReadGuestMetricsRequest,
+        },
+    },
+};
+use log::error;
+use std::time::Duration;
+use tokio::{
+    select,
+    sync::mpsc::{channel, Receiver, Sender},
+    task::JoinHandle,
+    time::{sleep, timeout},
+};
+use tonic::transport::Channel;
+
+use crate::format::metrics_value_pretty;
+
+pub struct MetricState {
+    pub guest: Guest,
+    pub root: Option<GuestMetricNode>,
+}
+
+pub struct MultiMetricState {
+    pub guests: Vec<MetricState>,
+}
+
+pub struct MultiMetricCollector {
+    client: ControlServiceClient<Channel>,
+    events: EventStream,
+    period: Duration,
+}
+
+pub struct MultiMetricCollectorHandle {
+    pub receiver: Receiver<MultiMetricState>,
+    task: JoinHandle<()>,
+}
+
+impl Drop for MultiMetricCollectorHandle {
+    fn drop(&mut self) {
+        self.task.abort();
+    }
+}
+
+impl MultiMetricCollector {
+    pub fn new(
+        client: ControlServiceClient<Channel>,
+        events: EventStream,
+        period: Duration,
+    ) -> Result<MultiMetricCollector> {
+        Ok(MultiMetricCollector {
+            client,
+            events,
+            period,
+        })
+    }
+
+    pub async fn launch(mut self) -> Result<MultiMetricCollectorHandle> {
+        let (sender, receiver) = channel::<MultiMetricState>(100);
+        let task = tokio::task::spawn(async move {
+            if let Err(error) = self.process(sender).await {
+                error!("failed to process multi metric collector: {}", error);
+            }
+        });
+        Ok(MultiMetricCollectorHandle { receiver, task })
+    }
+
+    pub async fn process(&mut self, sender: Sender<MultiMetricState>) -> Result<()> {
+        let mut events = self.events.subscribe();
+        let mut guests: Vec<Guest> = self
+            .client
+            .list_guests(ListGuestsRequest {})
+            .await?
+            .into_inner()
+            .guests;
+        loop {
+            let collect = select! {
+                x = events.recv() => match x {
+                    Ok(event) => {
+                        let Event::GuestChanged(changed) = event;
+                            let Some(guest) = changed.guest else {
+                                continue;
+                            };
+                            let Some(ref state) = guest.state else {
+                                continue;
+                            };
+                            guests.retain(|x| x.id != guest.id);
+                            if state.status() != GuestStatus::Destroying {
+                                guests.push(guest);
+                            }
+                        false
+                    },
+
+                    Err(error) => {
+                        return Err(error.into());
+                    }
+                },
+
+                _ = sleep(self.period) => {
+                    true
+                }
+            };
+
+            if !collect {
+                continue;
+            }
+
+            let mut metrics = Vec::new();
+            for guest in &guests {
+                let Some(ref state) = guest.state else {
+                    continue;
+                };
+
+                if state.status() != GuestStatus::Started {
+                    continue;
+                }
+
+                let root = timeout(
+                    Duration::from_secs(5),
+                    self.client.read_guest_metrics(ReadGuestMetricsRequest {
+                        guest_id: guest.id.clone(),
+                    }),
+                )
+                .await
+                .ok()
+                .and_then(|x| x.ok())
+                .map(|x| x.into_inner())
+                .and_then(|x| x.root);
+                metrics.push(MetricState {
+                    guest: guest.clone(),
+                    root,
+                });
+            }
+            sender.send(MultiMetricState { guests: metrics }).await?;
+        }
+    }
+}
+
+pub fn lookup<'a>(node: &'a GuestMetricNode, path: &str) -> Option<&'a GuestMetricNode> {
+    let Some((what, b)) = path.split_once('/') else {
+        return node.children.iter().find(|x| x.name == path);
+    };
+    let next = node.children.iter().find(|x| x.name == what)?;
+    return lookup(next, b);
+}
+
+pub fn lookup_metric_value(node: &GuestMetricNode, path: &str) -> Option<String> {
+    lookup(node, path).and_then(|x| {
+        x.value
+            .as_ref()
+            .map(|v| metrics_value_pretty(v.clone(), x.format()))
+    })
+}

crates/ctl/src/pull.rs

@@ -0,0 +1,268 @@
+use std::{
+    collections::{hash_map::Entry, HashMap},
+    time::Duration,
+};
+
+use anyhow::{anyhow, Result};
+use indicatif::{MultiProgress, ProgressBar, ProgressStyle};
+use krata::v1::control::{
+    image_progress_indication::Indication, ImageProgressIndication, ImageProgressLayerPhase,
+    ImageProgressPhase, PullImageReply,
+};
+use tokio_stream::StreamExt;
+use tonic::Streaming;
+
+const SPINNER_STRINGS: &[&str] = &[
+    "[=                   ]",
+    "[ =                  ]",
+    "[  =                 ]",
+    "[   =                ]",
+    "[    =               ]",
+    "[     =              ]",
+    "[      =             ]",
+    "[       =            ]",
+    "[        =           ]",
+    "[         =          ]",
+    "[          =         ]",
+    "[           =        ]",
+    "[            =       ]",
+    "[             =      ]",
+    "[              =     ]",
+    "[               =    ]",
+    "[                =   ]",
+    "[                 =  ]",
+    "[                  = ]",
+    "[                   =]",
+    "[====================]",
+];
+
+fn progress_bar_for_indication(indication: &ImageProgressIndication) -> Option<ProgressBar> {
+    match indication.indication.as_ref() {
+        Some(Indication::Hidden(_)) | None => None,
+        Some(Indication::Bar(indic)) => {
+            let bar = ProgressBar::new(indic.total);
+            bar.enable_steady_tick(Duration::from_millis(100));
+            Some(bar)
+        }
+        Some(Indication::Spinner(_)) => {
+            let bar = ProgressBar::new_spinner();
+            bar.enable_steady_tick(Duration::from_millis(100));
+            Some(bar)
+        }
+        Some(Indication::Completed(indic)) => {
+            let bar = ProgressBar::new_spinner();
+            bar.enable_steady_tick(Duration::from_millis(100));
+            if !indic.message.is_empty() {
+                bar.finish_with_message(indic.message.clone());
+            } else {
+                bar.finish()
+            }
+            Some(bar)
+        }
+    }
+}
+
+fn configure_for_indication(
+    bar: &mut ProgressBar,
+    multi_progress: &mut MultiProgress,
+    indication: &ImageProgressIndication,
+    top_phase: Option<ImageProgressPhase>,
+    layer_phase: Option<ImageProgressLayerPhase>,
+    layer_id: Option<&str>,
+) {
+    let prefix = if let Some(phase) = top_phase {
+        match phase {
+            ImageProgressPhase::Unknown => "unknown",
+            ImageProgressPhase::Started => "started",
+            ImageProgressPhase::Resolving => "resolving",
+            ImageProgressPhase::Resolved => "resolved",
+            ImageProgressPhase::ConfigDownload => "downloading",
+            ImageProgressPhase::LayerDownload => "downloading",
+            ImageProgressPhase::Assemble => "assembling",
+            ImageProgressPhase::Pack => "packing",
+            ImageProgressPhase::Complete => "complete",
+        }
+    } else if let Some(phase) = layer_phase {
+        match phase {
+            ImageProgressLayerPhase::Unknown => "unknown",
+            ImageProgressLayerPhase::Waiting => "waiting",
+            ImageProgressLayerPhase::Downloading => "downloading",
+            ImageProgressLayerPhase::Downloaded => "downloaded",
+            ImageProgressLayerPhase::Extracting => "extracting",
+            ImageProgressLayerPhase::Extracted => "extracted",
+        }
+    } else {
+        ""
+    };
+    let prefix = prefix.to_string();
+
+    let id = if let Some(layer_id) = layer_id {
+        let hash = if let Some((_, hash)) = layer_id.split_once(':') {
+            hash
+        } else {
+            "unknown"
+        };
+        let small_hash = if hash.len() > 10 { &hash[0..10] } else { hash };
+        Some(format!("{:width$}", small_hash, width = 10))
+    } else {
+        None
+    };
+
+    let prefix = if let Some(id) = id {
+        format!("{} {:width$}", id, prefix, width = 11)
+    } else {
+        format!("           {:width$}", prefix, width = 11)
+    };
+
+    match indication.indication.as_ref() {
+        Some(Indication::Hidden(_)) | None => {
+            multi_progress.remove(bar);
+            return;
+        }
+        Some(Indication::Bar(indic)) => {
+            if indic.is_bytes {
+                bar.set_style(ProgressStyle::with_template("{prefix} [{bar:20}] {msg} {binary_bytes}/{binary_total_bytes} ({binary_bytes_per_sec}) eta: {eta}").unwrap().progress_chars("=>-"));
+            } else {
+                bar.set_style(
+                    ProgressStyle::with_template(
+                        "{prefix} [{bar:20} {msg} {human_pos}/{human_len} ({per_sec}/sec)",
+                    )
+                    .unwrap()
+                    .progress_chars("=>-"),
+                );
+            }
+            bar.set_message(indic.message.clone());
+            bar.set_position(indic.current);
+            bar.set_length(indic.total);
+        }
+        Some(Indication::Spinner(indic)) => {
+            bar.set_style(
+                ProgressStyle::with_template("{prefix} {spinner}  {msg}")
+                    .unwrap()
+                    .tick_strings(SPINNER_STRINGS),
+            );
+            bar.set_message(indic.message.clone());
+        }
+        Some(Indication::Completed(indic)) => {
+            if bar.is_finished() {
+                return;
+            }
+            bar.disable_steady_tick();
+            bar.set_message(indic.message.clone());
+            if indic.total != 0 {
+                bar.set_position(indic.total);
+                bar.set_length(indic.total);
+            }
+            if bar.style().get_tick_str(0).contains('=') {
+                bar.set_style(
+                    ProgressStyle::with_template("{prefix} {spinner}  {msg}")
+                        .unwrap()
+                        .tick_strings(SPINNER_STRINGS),
+                );
+                bar.finish_with_message(indic.message.clone());
+            } else if indic.is_bytes {
+                bar.set_style(
+                    ProgressStyle::with_template("{prefix} [{bar:20}] {msg} {binary_total_bytes}")
+                        .unwrap()
+                        .progress_chars("=>-"),
+                );
+            } else {
+                bar.set_style(
+                    ProgressStyle::with_template("{prefix} [{bar:20}] {msg}")
+                        .unwrap()
+                        .progress_chars("=>-"),
+                );
+            }
+            bar.tick();
+            bar.enable_steady_tick(Duration::from_millis(100));
+        }
+    };
+
+    bar.set_prefix(prefix);
+    bar.tick();
+}
+
+pub async fn pull_interactive_progress(
+    mut stream: Streaming<PullImageReply>,
+) -> Result<PullImageReply> {
+    let mut multi_progress = MultiProgress::new();
+    multi_progress.set_move_cursor(false);
+    let mut progresses = HashMap::new();
+
+    while let Some(reply) = stream.next().await {
+        let reply = match reply {
+            Ok(reply) => reply,
+            Err(error) => {
+                multi_progress.clear()?;
+                return Err(error.into());
+            }
+        };
+
+        if reply.progress.is_none() && !reply.digest.is_empty() {
+            multi_progress.clear()?;
+            return Ok(reply);
+        }
+
+        let Some(oci) = reply.progress else {
+            continue;
+        };
+
+        for layer in &oci.layers {
+            let Some(ref indication) = layer.indication else {
+                continue;
+            };
+
+            let bar = match progresses.entry(layer.id.clone()) {
+                Entry::Occupied(entry) => Some(entry.into_mut()),
+
+                Entry::Vacant(entry) => {
+                    if let Some(bar) = progress_bar_for_indication(indication) {
+                        multi_progress.add(bar.clone());
+                        Some(entry.insert(bar))
+                    } else {
+                        None
+                    }
+                }
+            };
+
+            if let Some(bar) = bar {
+                configure_for_indication(
+                    bar,
+                    &mut multi_progress,
+                    indication,
+                    None,
+                    Some(layer.phase()),
+                    Some(&layer.id),
+                );
+            }
+        }
+
+        if let Some(ref indication) = oci.indication {
+            let bar = match progresses.entry("root".to_string()) {
+                Entry::Occupied(entry) => Some(entry.into_mut()),
+
+                Entry::Vacant(entry) => {
+                    if let Some(bar) = progress_bar_for_indication(indication) {
+                        multi_progress.add(bar.clone());
+                        Some(entry.insert(bar))
+                    } else {
+                        None
+                    }
+                }
+            };
+
+            if let Some(bar) = bar {
+                configure_for_indication(
+                    bar,
+                    &mut multi_progress,
+                    indication,
+                    Some(oci.phase()),
+                    None,
+                    None,
+                );
+            }
+        }
+    }
+    multi_progress.clear()?;
+    Err(anyhow!("never received final reply for image pull"))
+}

crates/daemon/Cargo.toml

@@ -2,7 +2,7 @@
 name = "krata-daemon"
 description = "Daemon for the krata hypervisor."
 license.workspace = true
-version= "0.0.3"
+version.workspace = true
 homepage.workspace = true
 repository.workspace = true
 edition = "2021"
@@ -13,17 +13,21 @@ anyhow = { workspace = true }
 async-stream = { workspace = true }
 async-trait = { workspace = true }
 bytes = { workspace = true }
+circular-buffer = { workspace = true }
 clap = { workspace = true }
 env_logger = { workspace = true }
 futures = { workspace = true }
-krata = { path = "../krata", version = "^0.0.3" }
-krata-runtime = { path = "../runtime", version = "^0.0.3" }
+krata = { path = "../krata", version = "^0.0.10" }
+krata-oci = { path = "../oci", version = "^0.0.10" }
+krata-runtime = { path = "../runtime", version = "^0.0.10" }
 log = { workspace = true }
 prost = { workspace = true }
 redb = { workspace = true }
+scopeguard = { workspace = true }
 signal-hook = { workspace = true }
 tokio = { workspace = true }
 tokio-stream = { workspace = true }
+krata-tokio-tar = { workspace = true }
 tonic = { workspace = true, features = ["tls"] }
 uuid = { workspace = true }
 

crates/daemon/bin/daemon.rs

@@ -1,22 +1,9 @@
 use anyhow::Result;
 use clap::Parser;
 use env_logger::Env;
-use krata::dial::ControlDialAddress;
-use kratad::Daemon;
-use kratart::Runtime;
+use kratad::command::DaemonCommand;
 use log::LevelFilter;
-use std::{
-    str::FromStr,
-    sync::{atomic::AtomicBool, Arc},
-};
-
-#[derive(Parser)]
-struct DaemonCommand {
-    #[arg(short, long, default_value = "unix:///var/lib/krata/daemon.socket")]
-    listen: String,
-    #[arg(short, long, default_value = "/var/lib/krata")]
-    store: String,
-}
+use std::sync::{atomic::AtomicBool, Arc};
 
 #[tokio::main(flavor = "multi_thread", worker_threads = 10)]
 async fn main() -> Result<()> {
@@ -25,12 +12,8 @@ async fn main() -> Result<()> {
         .init();
     mask_sighup()?;
 
-    let args = DaemonCommand::parse();
-    let addr = ControlDialAddress::from_str(&args.listen)?;
-    let runtime = Runtime::new(args.store.clone()).await?;
-    let mut daemon = Daemon::new(args.store.clone(), runtime).await?;
-    daemon.listen(addr).await?;
-    Ok(())
+    let command = DaemonCommand::parse();
+    command.run().await
 }
 
 fn mask_sighup() -> Result<()> {

crates/daemon/src/command.rs

@@ -0,0 +1,36 @@
+use anyhow::Result;
+use clap::{CommandFactory, Parser};
+use krata::dial::ControlDialAddress;
+use std::str::FromStr;
+
+use crate::Daemon;
+
+#[derive(Parser)]
+#[command(version, about = "Krata hypervisor daemon")]
+pub struct DaemonCommand {
+    #[arg(
+        short,
+        long,
+        default_value = "unix:///var/lib/krata/daemon.socket",
+        help = "Listen address"
+    )]
+    listen: String,
+    #[arg(short, long, default_value = "/var/lib/krata", help = "Storage path")]
+    store: String,
+}
+
+impl DaemonCommand {
+    pub async fn run(self) -> Result<()> {
+        let addr = ControlDialAddress::from_str(&self.listen)?;
+        let mut daemon = Daemon::new(self.store.clone()).await?;
+        daemon.listen(addr).await?;
+        Ok(())
+    }
+
+    pub fn version() -> String {
+        DaemonCommand::command()
+            .get_version()
+            .unwrap_or("unknown")
+            .to_string()
+    }
+}

crates/daemon/src/console.rs

@@ -0,0 +1,164 @@
+use std::{collections::HashMap, sync::Arc};
+
+use anyhow::{anyhow, Result};
+use circular_buffer::CircularBuffer;
+use kratart::channel::ChannelService;
+use log::error;
+use tokio::{
+    sync::{
+        mpsc::{error::TrySendError, Receiver, Sender},
+        Mutex,
+    },
+    task::JoinHandle,
+};
+use uuid::Uuid;
+
+use crate::glt::GuestLookupTable;
+
+const CONSOLE_BUFFER_SIZE: usize = 1024 * 1024;
+type RawConsoleBuffer = CircularBuffer<CONSOLE_BUFFER_SIZE, u8>;
+type ConsoleBuffer = Box<RawConsoleBuffer>;
+
+type ListenerMap = Arc<Mutex<HashMap<u32, Vec<Sender<Vec<u8>>>>>>;
+type BufferMap = Arc<Mutex<HashMap<u32, ConsoleBuffer>>>;
+
+#[derive(Clone)]
+pub struct DaemonConsoleHandle {
+    glt: GuestLookupTable,
+    listeners: ListenerMap,
+    buffers: BufferMap,
+    sender: Sender<(u32, Vec<u8>)>,
+    task: Arc<JoinHandle<()>>,
+}
+
+#[derive(Clone)]
+pub struct DaemonConsoleAttachHandle {
+    pub initial: Vec<u8>,
+    listeners: ListenerMap,
+    sender: Sender<(u32, Vec<u8>)>,
+    domid: u32,
+}
+
+impl DaemonConsoleAttachHandle {
+    pub async fn unsubscribe(&self) -> Result<()> {
+        let mut guard = self.listeners.lock().await;
+        let _ = guard.remove(&self.domid);
+        Ok(())
+    }
+
+    pub async fn send(&self, data: Vec<u8>) -> Result<()> {
+        Ok(self.sender.send((self.domid, data)).await?)
+    }
+}
+
+impl DaemonConsoleHandle {
+    pub async fn attach(
+        &self,
+        uuid: Uuid,
+        sender: Sender<Vec<u8>>,
+    ) -> Result<DaemonConsoleAttachHandle> {
+        let Some(domid) = self.glt.lookup_domid_by_uuid(&uuid).await else {
+            return Err(anyhow!("unable to find domain {}", uuid));
+        };
+        let buffers = self.buffers.lock().await;
+        let buffer = buffers.get(&domid).map(|x| x.to_vec()).unwrap_or_default();
+        drop(buffers);
+        let mut listeners = self.listeners.lock().await;
+        let senders = listeners.entry(domid).or_default();
+        senders.push(sender);
+        Ok(DaemonConsoleAttachHandle {
+            initial: buffer,
+            sender: self.sender.clone(),
+            listeners: self.listeners.clone(),
+            domid,
+        })
+    }
+}
+
+impl Drop for DaemonConsoleHandle {
+    fn drop(&mut self) {
+        if Arc::strong_count(&self.task) <= 1 {
+            self.task.abort();
+        }
+    }
+}
+
+pub struct DaemonConsole {
+    glt: GuestLookupTable,
+    listeners: ListenerMap,
+    buffers: BufferMap,
+    receiver: Receiver<(u32, Option<Vec<u8>>)>,
+    sender: Sender<(u32, Vec<u8>)>,
+    task: JoinHandle<()>,
+}
+
+impl DaemonConsole {
+    pub async fn new(glt: GuestLookupTable) -> Result<DaemonConsole> {
+        let (service, sender, receiver) =
+            ChannelService::new("krata-console".to_string(), Some(0)).await?;
+        let task = service.launch().await?;
+        let listeners = Arc::new(Mutex::new(HashMap::new()));
+        let buffers = Arc::new(Mutex::new(HashMap::new()));
+        Ok(DaemonConsole {
+            glt,
+            listeners,
+            buffers,
+            receiver,
+            sender,
+            task,
+        })
+    }
+
+    pub async fn launch(mut self) -> Result<DaemonConsoleHandle> {
+        let glt = self.glt.clone();
+        let listeners = self.listeners.clone();
+        let buffers = self.buffers.clone();
+        let sender = self.sender.clone();
+        let task = tokio::task::spawn(async move {
+            if let Err(error) = self.process().await {
+                error!("failed to process console: {}", error);
+            }
+        });
+        Ok(DaemonConsoleHandle {
+            glt,
+            listeners,
+            buffers,
+            sender,
+            task: Arc::new(task),
+        })
+    }
+
+    async fn process(&mut self) -> Result<()> {
+        loop {
+            let Some((domid, data)) = self.receiver.recv().await else {
+                break;
+            };
+
+            let mut buffers = self.buffers.lock().await;
+            if let Some(data) = data {
+                let buffer = buffers
+                    .entry(domid)
+                    .or_insert_with_key(|_| RawConsoleBuffer::boxed());
+                buffer.extend_from_slice(&data);
+                drop(buffers);
+                let mut listeners = self.listeners.lock().await;
+                if let Some(senders) = listeners.get_mut(&domid) {
+                    senders.retain(|sender| {
+                        !matches!(sender.try_send(data.to_vec()), Err(TrySendError::Closed(_)))
+                    });
+                }
+            } else {
+                buffers.remove(&domid);
+                let mut listeners = self.listeners.lock().await;
+                listeners.remove(&domid);
+            }
+        }
+        Ok(())
+    }
+}
+
+impl Drop for DaemonConsole {
+    fn drop(&mut self) {
+        self.task.abort();
+    }
+}

crates/daemon/src/control.rs

@@ -1,27 +1,43 @@
-use std::{io, pin::Pin, str::FromStr};
-
 use async_stream::try_stream;
 use futures::Stream;
-use krata::v1::{
-    common::{Guest, GuestState, GuestStatus},
-    control::{
-        control_service_server::ControlService, ConsoleDataReply, ConsoleDataRequest,
-        CreateGuestReply, CreateGuestRequest, DestroyGuestReply, DestroyGuestRequest,
-        ListGuestsReply, ListGuestsRequest, ResolveGuestReply, ResolveGuestRequest,
-        WatchEventsReply, WatchEventsRequest,
+use krata::{
+    idm::internal::{
+        exec_stream_request_update::Update, request::Request as IdmRequestType,
+        response::Response as IdmResponseType, ExecEnvVar, ExecStreamRequestStart,
+        ExecStreamRequestStdin, ExecStreamRequestUpdate, MetricsRequest, Request as IdmRequest,
+    },
+    v1::{
+        common::{Guest, GuestState, GuestStatus, OciImageFormat},
+        control::{
+            control_service_server::ControlService, ConsoleDataReply, ConsoleDataRequest,
+            CreateGuestReply, CreateGuestRequest, DestroyGuestReply, DestroyGuestRequest,
+            ExecGuestReply, ExecGuestRequest, IdentifyHostReply, IdentifyHostRequest,
+            ListGuestsReply, ListGuestsRequest, PullImageReply, PullImageRequest,
+            ReadGuestMetricsReply, ReadGuestMetricsRequest, ResolveGuestReply, ResolveGuestRequest,
+            SnoopIdmReply, SnoopIdmRequest, WatchEventsReply, WatchEventsRequest,
+        },
     },
 };
-use kratart::Runtime;
+use krataoci::{
+    name::ImageName,
+    packer::{service::OciPackerService, OciPackedFormat, OciPackedImage},
+    progress::{OciProgress, OciProgressContext},
+};
+use std::{pin::Pin, str::FromStr};
 use tokio::{
-    io::{AsyncReadExt, AsyncWriteExt},
     select,
-    sync::mpsc::Sender,
+    sync::mpsc::{channel, Sender},
+    task::JoinError,
 };
 use tokio_stream::StreamExt;
 use tonic::{Request, Response, Status, Streaming};
 use uuid::Uuid;
 
-use crate::{db::GuestStore, event::DaemonEventContext};
+use crate::{
+    command::DaemonCommand, console::DaemonConsoleHandle, db::GuestStore,
+    event::DaemonEventContext, glt::GuestLookupTable, idm::DaemonIdmHandle,
+    metrics::idm_metric_to_api, oci::convert_oci_progress,
+};
 
 pub struct ApiError {
     message: String,
@@ -42,42 +58,77 @@ impl From<ApiError> for Status {
 }
 
 #[derive(Clone)]
-pub struct RuntimeControlService {
+pub struct DaemonControlService {
+    glt: GuestLookupTable,
     events: DaemonEventContext,
-    runtime: Runtime,
+    console: DaemonConsoleHandle,
+    idm: DaemonIdmHandle,
     guests: GuestStore,
     guest_reconciler_notify: Sender<Uuid>,
+    packer: OciPackerService,
 }
 
-impl RuntimeControlService {
+impl DaemonControlService {
     pub fn new(
+        glt: GuestLookupTable,
         events: DaemonEventContext,
-        runtime: Runtime,
+        console: DaemonConsoleHandle,
+        idm: DaemonIdmHandle,
         guests: GuestStore,
         guest_reconciler_notify: Sender<Uuid>,
+        packer: OciPackerService,
     ) -> Self {
         Self {
+            glt,
             events,
-            runtime,
+            console,
+            idm,
             guests,
             guest_reconciler_notify,
+            packer,
         }
     }
 }
 
 enum ConsoleDataSelect {
-    Read(io::Result<usize>),
+    Read(Option<Vec<u8>>),
     Write(Option<Result<ConsoleDataRequest, tonic::Status>>),
 }
 
+enum PullImageSelect {
+    Progress(Option<OciProgress>),
+    Completed(Result<Result<OciPackedImage, anyhow::Error>, JoinError>),
+}
+
 #[tonic::async_trait]
-impl ControlService for RuntimeControlService {
+impl ControlService for DaemonControlService {
+    type ExecGuestStream =
+        Pin<Box<dyn Stream<Item = Result<ExecGuestReply, Status>> + Send + 'static>>;
+
     type ConsoleDataStream =
         Pin<Box<dyn Stream<Item = Result<ConsoleDataReply, Status>> + Send + 'static>>;
 
+    type PullImageStream =
+        Pin<Box<dyn Stream<Item = Result<PullImageReply, Status>> + Send + 'static>>;
+
     type WatchEventsStream =
         Pin<Box<dyn Stream<Item = Result<WatchEventsReply, Status>> + Send + 'static>>;
 
+    type SnoopIdmStream =
+        Pin<Box<dyn Stream<Item = Result<SnoopIdmReply, Status>> + Send + 'static>>;
+
+    async fn identify_host(
+        &self,
+        request: Request<IdentifyHostRequest>,
+    ) -> Result<Response<IdentifyHostReply>, Status> {
+        let _ = request.into_inner();
+        Ok(Response::new(IdentifyHostReply {
+            host_domid: self.glt.host_domid(),
+            host_uuid: self.glt.host_uuid().to_string(),
+            krata_version: DaemonCommand::version(),
+        }))
+    }
+
     async fn create_guest(
         &self,
         request: Request<CreateGuestRequest>,
@@ -100,6 +151,7 @@ impl ControlService for RuntimeControlService {
                         network: None,
                         exit_info: None,
                         error_info: None,
+                        host: self.glt.host_uuid().to_string(),
                         domid: u32::MAX,
                     }),
                     spec: Some(spec),
@@ -118,6 +170,98 @@ impl ControlService for RuntimeControlService {
         }))
     }
 
+    async fn exec_guest(
+        &self,
+        request: Request<Streaming<ExecGuestRequest>>,
+    ) -> Result<Response<Self::ExecGuestStream>, Status> {
+        let mut input = request.into_inner();
+        let Some(request) = input.next().await else {
+            return Err(ApiError {
+                message: "expected to have at least one request".to_string(),
+            }
+            .into());
+        };
+        let request = request?;
+
+        let Some(task) = request.task else {
+            return Err(ApiError {
+                message: "task is missing".to_string(),
+            }
+            .into());
+        };
+
+        let uuid = Uuid::from_str(&request.guest_id).map_err(|error| ApiError {
+            message: error.to_string(),
+        })?;
+        let idm = self.idm.client(uuid).await.map_err(|error| ApiError {
+            message: error.to_string(),
+        })?;
+
+        let idm_request = IdmRequest {
+            request: Some(IdmRequestType::ExecStream(ExecStreamRequestUpdate {
+                update: Some(Update::Start(ExecStreamRequestStart {
+                    environment: task
+                        .environment
+                        .into_iter()
+                        .map(|x| ExecEnvVar {
+                            key: x.key,
+                            value: x.value,
+                        })
+                        .collect(),
+                    command: task.command,
+                    working_directory: task.working_directory,
+                })),
+            })),
+        };
+
+        let output = try_stream! {
+            let mut handle = idm.send_stream(idm_request).await.map_err(|x| ApiError {
+                message: x.to_string(),
+            })?;
+
+            loop {
+                select! {
+                    x = input.next() => if let Some(update) = x {
+                        let update: Result<ExecGuestRequest, Status> = update.map_err(|error| ApiError {
+                            message: error.to_string()
+                        }.into());
+
+                        if let Ok(update) = update {
+                            if !update.data.is_empty() {
+                                let _ = handle.update(IdmRequest {
+                                    request: Some(IdmRequestType::ExecStream(ExecStreamRequestUpdate {
+                                        update: Some(Update::Stdin(ExecStreamRequestStdin {
+                                            data: update.data,
+                                        })),
+                                    }))}).await;
+                            }
+                        }
+                    },
+                    x = handle.receiver.recv() => match x {
+                        Some(response) => {
+                            let Some(IdmResponseType::ExecStream(update)) = response.response else {
+                                break;
+                            };
+                            let reply = ExecGuestReply {
+                                exited: update.exited,
+                                error: update.error,
+                                exit_code: update.exit_code,
+                                stdout: update.stdout,
+                                stderr: update.stderr
+                            };
+                            yield reply;
+                        },
+                        None => {
+                            break;
+                        }
+                    }
+                };
+            }
+        };
+
+        Ok(Response::new(Box::pin(output) as Self::ExecGuestStream))
+    }
+
     async fn destroy_guest(
         &self,
         request: Request<DestroyGuestRequest>,
@@ -200,27 +344,38 @@ impl ControlService for RuntimeControlService {
         let uuid = Uuid::from_str(&request.guest_id).map_err(|error| ApiError {
             message: error.to_string(),
         })?;
-        let mut console = self.runtime.console(uuid).await.map_err(ApiError::from)?;
+        let (sender, mut receiver) = channel(100);
+        let console = self
+            .console
+            .attach(uuid, sender)
+            .await
+            .map_err(|error| ApiError {
+                message: format!("failed to attach to console: {}", error),
+            })?;
 
         let output = try_stream! {
-            let mut buffer: Vec<u8> = vec![0u8; 256];
+            yield ConsoleDataReply { data: console.initial.clone(), };
             loop {
                 let what = select! {
-                    x = console.read_handle.read(&mut buffer) => ConsoleDataSelect::Read(x),
+                    x = receiver.recv() => ConsoleDataSelect::Read(x),
                     x = input.next() => ConsoleDataSelect::Write(x),
                 };
 
                 match what {
-                    ConsoleDataSelect::Read(result) => {
-                        let size = result?;
-                        let data = buffer[0..size].to_vec();
+                    ConsoleDataSelect::Read(Some(data)) => {
                         yield ConsoleDataReply { data, };
                     },
 
+                    ConsoleDataSelect::Read(None) => {
+                        break;
+                    }
+
                     ConsoleDataSelect::Write(Some(request)) => {
                         let request = request?;
                         if !request.data.is_empty() {
-                            console.write_handle.write_all(&request.data).await?;
+                            console.send(request.data).await.map_err(|error| ApiError {
+                                message: error.to_string(),
+                            })?;
                         }
                     },
 
@@ -234,6 +389,107 @@ impl ControlService for RuntimeControlService {
         Ok(Response::new(Box::pin(output) as Self::ConsoleDataStream))
     }
 
+    async fn read_guest_metrics(
+        &self,
+        request: Request<ReadGuestMetricsRequest>,
+    ) -> Result<Response<ReadGuestMetricsReply>, Status> {
+        let request = request.into_inner();
+        let uuid = Uuid::from_str(&request.guest_id).map_err(|error| ApiError {
+            message: error.to_string(),
+        })?;
+        let client = self.idm.client(uuid).await.map_err(|error| ApiError {
+            message: error.to_string(),
+        })?;
+
+        let response = client
+            .send(IdmRequest {
+                request: Some(IdmRequestType::Metrics(MetricsRequest {})),
+            })
+            .await
+            .map_err(|error| ApiError {
+                message: error.to_string(),
+            })?;
+
+        let mut reply = ReadGuestMetricsReply::default();
+        if let Some(IdmResponseType::Metrics(metrics)) = response.response {
+            reply.root = metrics.root.map(idm_metric_to_api);
+        }
+        Ok(Response::new(reply))
+    }
+
+    async fn pull_image(
+        &self,
+        request: Request<PullImageRequest>,
+    ) -> Result<Response<Self::PullImageStream>, Status> {
+        let request = request.into_inner();
+        let name = ImageName::parse(&request.image).map_err(|err| ApiError {
+            message: err.to_string(),
+        })?;
+        let format = match request.format() {
+            OciImageFormat::Unknown => OciPackedFormat::Squashfs,
+            OciImageFormat::Squashfs => OciPackedFormat::Squashfs,
+            OciImageFormat::Erofs => OciPackedFormat::Erofs,
+            OciImageFormat::Tar => OciPackedFormat::Tar,
+        };
+        let (context, mut receiver) = OciProgressContext::create();
+        let our_packer = self.packer.clone();
+
+        let output = try_stream! {
+            let mut task = tokio::task::spawn(async move {
+                our_packer.request(name, format, request.overwrite_cache, context).await
+            });
+            let abort_handle = task.abort_handle();
+            let _task_cancel_guard = scopeguard::guard(abort_handle, |handle| {
+                handle.abort();
+            });
+
+            loop {
+                let what = select! {
+                    x = receiver.changed() => match x {
+                        Ok(_) => PullImageSelect::Progress(Some(receiver.borrow_and_update().clone())),
+                        Err(_) => PullImageSelect::Progress(None),
+                    },
+                    x = &mut task => PullImageSelect::Completed(x),
+                };
+                match what {
+                    PullImageSelect::Progress(Some(progress)) => {
+                        let reply = PullImageReply {
+                            progress: Some(convert_oci_progress(progress)),
+                            digest: String::new(),
+                            format: OciImageFormat::Unknown.into(),
+                        };
+                        yield reply;
+                    },
+
+                    PullImageSelect::Completed(result) => {
+                        let result = result.map_err(|err| ApiError {
+                            message: err.to_string(),
+                        })?;
+                        let packed = result.map_err(|err| ApiError {
+                            message: err.to_string(),
+                        })?;
+                        let reply = PullImageReply {
+                            progress: None,
+                            digest: packed.digest,
+                            format: match packed.format {
+                                OciPackedFormat::Squashfs => OciImageFormat::Squashfs.into(),
+                                OciPackedFormat::Erofs => OciImageFormat::Erofs.into(),
+                                OciPackedFormat::Tar => OciImageFormat::Tar.into(),
+                            },
+                        };
+                        yield reply;
+                        break;
+                    },
+
+                    _ => {
+                        continue;
+                    }
+                }
+            }
+        };
+        Ok(Response::new(Box::pin(output) as Self::PullImageStream))
+    }
+
     async fn watch_events(
         &self,
         request: Request<WatchEventsRequest>,
@@ -247,4 +503,25 @@ impl ControlService for RuntimeControlService {
         };
         Ok(Response::new(Box::pin(output) as Self::WatchEventsStream))
     }
+
+    async fn snoop_idm(
+        &self,
+        request: Request<SnoopIdmRequest>,
+    ) -> Result<Response<Self::SnoopIdmStream>, Status> {
+        let _ = request.into_inner();
+        let mut messages = self.idm.snoop();
+        let glt = self.glt.clone();
+        let output = try_stream! {
+            while let Ok(event) = messages.recv().await {
+                let Some(from_uuid) = glt.lookup_uuid_by_domid(event.from).await else {
+                    continue;
+                };
+                let Some(to_uuid) = glt.lookup_uuid_by_domid(event.to).await else {
+                    continue;
+                };
+                yield SnoopIdmReply { from: from_uuid.to_string(), to: to_uuid.to_string(), packet: Some(event.packet) };
+            }
+        };
+        Ok(Response::new(Box::pin(output) as Self::SnoopIdmStream))
+    }
 }

crates/daemon/src/event.rs

@@ -6,10 +6,10 @@ use std::{
 
 use anyhow::Result;
 use krata::{
-    idm::protocol::{idm_event::Event, IdmPacket},
+    idm::{internal::event::Event as EventType, internal::Event},
     v1::common::{GuestExitInfo, GuestState, GuestStatus},
 };
-use log::error;
+use log::{error, warn};
 use tokio::{
     select,
     sync::{
@@ -21,15 +21,12 @@ use tokio::{
 };
 use uuid::Uuid;
 
-use crate::{
-    db::GuestStore,
-    idm::{DaemonIdmHandle, DaemonIdmSubscribeHandle},
-};
+use crate::{db::GuestStore, idm::DaemonIdmHandle};
 
 pub type DaemonEvent = krata::v1::control::watch_events_reply::Event;
 
 const EVENT_CHANNEL_QUEUE_LEN: usize = 1000;
-const IDM_CHANNEL_QUEUE_LEN: usize = 1000;
+const IDM_EVENT_CHANNEL_QUEUE_LEN: usize = 1000;
 
 #[derive(Clone)]
 pub struct DaemonEventContext {
@@ -52,9 +49,9 @@ pub struct DaemonEventGenerator {
     guest_reconciler_notify: Sender<Uuid>,
     feed: broadcast::Receiver<DaemonEvent>,
     idm: DaemonIdmHandle,
-    idms: HashMap<u32, (Uuid, DaemonIdmSubscribeHandle)>,
-    idm_sender: Sender<(u32, IdmPacket)>,
-    idm_receiver: Receiver<(u32, IdmPacket)>,
+    idms: HashMap<u32, (Uuid, JoinHandle<()>)>,
+    idm_sender: Sender<(u32, Event)>,
+    idm_receiver: Receiver<(u32, Event)>,
     _event_sender: broadcast::Sender<DaemonEvent>,
 }
 
@@ -65,7 +62,7 @@ impl DaemonEventGenerator {
         idm: DaemonIdmHandle,
     ) -> Result<(DaemonEventContext, DaemonEventGenerator)> {
         let (sender, _) = broadcast::channel(EVENT_CHANNEL_QUEUE_LEN);
-        let (idm_sender, idm_receiver) = channel(IDM_CHANNEL_QUEUE_LEN);
+        let (idm_sender, idm_receiver) = channel(IDM_EVENT_CHANNEL_QUEUE_LEN);
         let generator = DaemonEventGenerator {
             guests,
             guest_reconciler_notify,
@@ -81,46 +78,55 @@ impl DaemonEventGenerator {
     }
 
     async fn handle_feed_event(&mut self, event: &DaemonEvent) -> Result<()> {
-        match event {
-            DaemonEvent::GuestChanged(changed) => {
-                let Some(ref guest) = changed.guest else {
-                    return Ok(());
-                };
+        let DaemonEvent::GuestChanged(changed) = event;
+        let Some(ref guest) = changed.guest else {
+            return Ok(());
+        };
 
-                let Some(ref state) = guest.state else {
-                    return Ok(());
-                };
+        let Some(ref state) = guest.state else {
+            return Ok(());
+        };
 
-                let status = state.status();
-                let id = Uuid::from_str(&guest.id)?;
-                let domid = state.domid;
-                match status {
-                    GuestStatus::Started => {
-                        if let Entry::Vacant(e) = self.idms.entry(domid) {
-                            let subscribe =
-                                self.idm.subscribe(domid, self.idm_sender.clone()).await?;
-                            e.insert((id, subscribe));
+        let status = state.status();
+        let id = Uuid::from_str(&guest.id)?;
+        let domid = state.domid;
+        match status {
+            GuestStatus::Started => {
+                if let Entry::Vacant(e) = self.idms.entry(domid) {
+                    let client = self.idm.client_by_domid(domid).await?;
+                    let mut receiver = client.subscribe().await?;
+                    let sender = self.idm_sender.clone();
+                    let task = tokio::task::spawn(async move {
+                        loop {
+                            let Ok(event) = receiver.recv().await else {
+                                break;
+                            };
+
+                            if let Err(error) = sender.send((domid, event)).await {
+                                warn!("unable to deliver idm event: {}", error);
+                            }
                         }
-                    }
-
-                    GuestStatus::Destroyed => {
-                        if let Some((_, handle)) = self.idms.remove(&domid) {
-                            handle.unsubscribe().await?;
-                        }
-                    }
-
-                    _ => {}
+                    });
+                    e.insert((id, task));
                 }
             }
+
+            GuestStatus::Destroyed => {
+                if let Some((_, handle)) = self.idms.remove(&domid) {
+                    handle.abort();
+                }
+            }
+
+            _ => {}
         }
         Ok(())
     }
 
-    async fn handle_idm_packet(&mut self, id: Uuid, packet: IdmPacket) -> Result<()> {
-        if let Some(Event::Exit(exit)) = packet.event.and_then(|x| x.event) {
-            self.handle_exit_code(id, exit.code).await?;
+    async fn handle_idm_event(&mut self, id: Uuid, event: Event) -> Result<()> {
+        match event.event {
+            Some(EventType::Exit(exit)) => self.handle_exit_code(id, exit.code).await,
+            None => Ok(()),
         }
-        Ok(())
     }
 
     async fn handle_exit_code(&mut self, id: Uuid, code: i32) -> Result<()> {
@@ -130,6 +136,7 @@ impl DaemonEventGenerator {
                 network: guest.state.clone().unwrap_or_default().network,
                 exit_info: Some(GuestExitInfo { code }),
                 error_info: None,
+                host: guest.state.clone().map(|x| x.host).unwrap_or_default(),
                 domid: guest.state.clone().map(|x| x.domid).unwrap_or(u32::MAX),
             });
 
@@ -142,9 +149,9 @@ impl DaemonEventGenerator {
     async fn evaluate(&mut self) -> Result<()> {
         select! {
             x = self.idm_receiver.recv() => match x {
-                Some((domid, packet)) => {
+                Some((domid, event)) => {
                     if let Some((id, _)) = self.idms.get(&domid) {
-                        self.handle_idm_packet(*id, packet).await?;
+                        self.handle_idm_event(*id, event).await?;
                     }
                     Ok(())
                 },
@@ -159,7 +166,7 @@ impl DaemonEventGenerator {
                 Err(error) => {
                     Err(error.into())
                 }
-            }
+            },
         }
     }
 

crates/daemon/src/glt.rs

@@ -0,0 +1,69 @@
+use std::{collections::HashMap, sync::Arc};
+
+use tokio::sync::RwLock;
+use uuid::Uuid;
+
+struct GuestLookupTableState {
+    domid_to_uuid: HashMap<u32, Uuid>,
+    uuid_to_domid: HashMap<Uuid, u32>,
+}
+
+impl GuestLookupTableState {
+    pub fn new(host_uuid: Uuid) -> Self {
+        let mut domid_to_uuid = HashMap::new();
+        let mut uuid_to_domid = HashMap::new();
+        domid_to_uuid.insert(0, host_uuid);
+        uuid_to_domid.insert(host_uuid, 0);
+        GuestLookupTableState {
+            domid_to_uuid,
+            uuid_to_domid,
+        }
+    }
+}
+
+#[derive(Clone)]
+pub struct GuestLookupTable {
+    host_domid: u32,
+    host_uuid: Uuid,
+    state: Arc<RwLock<GuestLookupTableState>>,
+}
+
+impl GuestLookupTable {
+    pub fn new(host_domid: u32, host_uuid: Uuid) -> Self {
+        GuestLookupTable {
+            host_domid,
+            host_uuid,
+            state: Arc::new(RwLock::new(GuestLookupTableState::new(host_uuid))),
+        }
+    }
+
+    pub fn host_uuid(&self) -> Uuid {
+        self.host_uuid
+    }
+
+    pub fn host_domid(&self) -> u32 {
+        self.host_domid
+    }
+
+    pub async fn lookup_uuid_by_domid(&self, domid: u32) -> Option<Uuid> {
+        let state = self.state.read().await;
+        state.domid_to_uuid.get(&domid).cloned()
+    }
+
+    pub async fn lookup_domid_by_uuid(&self, uuid: &Uuid) -> Option<u32> {
+        let state = self.state.read().await;
+        state.uuid_to_domid.get(uuid).cloned()
+    }
+
+    pub async fn associate(&self, uuid: Uuid, domid: u32) {
+        let mut state = self.state.write().await;
+        state.uuid_to_domid.insert(uuid, domid);
+        state.domid_to_uuid.insert(domid, uuid);
+    }
+
+    pub async fn remove(&self, uuid: Uuid, domid: u32) {
+        let mut state = self.state.write().await;
+        state.uuid_to_domid.remove(&uuid);
+        state.domid_to_uuid.remove(&domid);
+    }
+}

crates/daemon/src/idm.rs

@@ -1,53 +1,58 @@
-use std::{collections::HashMap, sync::Arc};
+use std::{
+    collections::{hash_map::Entry, HashMap},
+    sync::Arc,
+};
 
-use anyhow::Result;
+use anyhow::{anyhow, Result};
 use bytes::{Buf, BytesMut};
-use krata::idm::protocol::IdmPacket;
+use krata::idm::{
+    client::{IdmBackend, IdmInternalClient},
+    internal::INTERNAL_IDM_CHANNEL,
+    transport::IdmTransportPacket,
+};
 use kratart::channel::ChannelService;
 use log::{error, warn};
 use prost::Message;
 use tokio::{
+    select,
     sync::{
-        mpsc::{Receiver, Sender},
+        broadcast,
+        mpsc::{channel, Receiver, Sender},
         Mutex,
     },
     task::JoinHandle,
 };
+use uuid::Uuid;
 
-type ListenerMap = Arc<Mutex<HashMap<u32, Sender<(u32, IdmPacket)>>>>;
+use crate::glt::GuestLookupTable;
+
+type BackendFeedMap = Arc<Mutex<HashMap<u32, Sender<IdmTransportPacket>>>>;
+type ClientMap = Arc<Mutex<HashMap<u32, IdmInternalClient>>>;
 
 #[derive(Clone)]
 pub struct DaemonIdmHandle {
-    listeners: ListenerMap,
+    glt: GuestLookupTable,
+    clients: ClientMap,
+    feeds: BackendFeedMap,
+    tx_sender: Sender<(u32, IdmTransportPacket)>,
     task: Arc<JoinHandle<()>>,
-}
-
-#[derive(Clone)]
-pub struct DaemonIdmSubscribeHandle {
-    domid: u32,
-    listeners: ListenerMap,
-}
-
-impl DaemonIdmSubscribeHandle {
-    pub async fn unsubscribe(&self) -> Result<()> {
-        let mut guard = self.listeners.lock().await;
-        let _ = guard.remove(&self.domid);
-        Ok(())
-    }
+    snoop_sender: broadcast::Sender<DaemonIdmSnoopPacket>,
 }
 
 impl DaemonIdmHandle {
-    pub async fn subscribe(
-        &self,
-        domid: u32,
-        sender: Sender<(u32, IdmPacket)>,
-    ) -> Result<DaemonIdmSubscribeHandle> {
-        let mut guard = self.listeners.lock().await;
-        guard.insert(domid, sender);
-        Ok(DaemonIdmSubscribeHandle {
-            domid,
-            listeners: self.listeners.clone(),
-        })
+    pub fn snoop(&self) -> broadcast::Receiver<DaemonIdmSnoopPacket> {
+        self.snoop_sender.subscribe()
+    }
+
+    pub async fn client(&self, uuid: Uuid) -> Result<IdmInternalClient> {
+        let Some(domid) = self.glt.lookup_domid_by_uuid(&uuid).await else {
+            return Err(anyhow!("unable to find domain {}", uuid));
+        };
+        self.client_by_domid(domid).await
+    }
+
+    pub async fn client_by_domid(&self, domid: u32) -> Result<IdmInternalClient> {
+        client_or_create(domid, &self.tx_sender, &self.clients, &self.feeds).await
     }
 }
 
@@ -59,70 +64,141 @@ impl Drop for DaemonIdmHandle {
     }
 }
 
+#[derive(Clone)]
+pub struct DaemonIdmSnoopPacket {
+    pub from: u32,
+    pub to: u32,
+    pub packet: IdmTransportPacket,
+}
+
 pub struct DaemonIdm {
-    listeners: ListenerMap,
-    receiver: Receiver<(u32, Vec<u8>)>,
+    glt: GuestLookupTable,
+    clients: ClientMap,
+    feeds: BackendFeedMap,
+    tx_sender: Sender<(u32, IdmTransportPacket)>,
+    tx_raw_sender: Sender<(u32, Vec<u8>)>,
+    tx_receiver: Receiver<(u32, IdmTransportPacket)>,
+    rx_receiver: Receiver<(u32, Option<Vec<u8>>)>,
+    snoop_sender: broadcast::Sender<DaemonIdmSnoopPacket>,
     task: JoinHandle<()>,
 }
 
 impl DaemonIdm {
-    pub async fn new() -> Result<DaemonIdm> {
-        let (service, receiver) = ChannelService::new("krata-channel".to_string()).await?;
+    pub async fn new(glt: GuestLookupTable) -> Result<DaemonIdm> {
+        let (service, tx_raw_sender, rx_receiver) =
+            ChannelService::new("krata-channel".to_string(), None).await?;
+        let (tx_sender, tx_receiver) = channel(100);
+        let (snoop_sender, _) = broadcast::channel(100);
         let task = service.launch().await?;
-        let listeners = Arc::new(Mutex::new(HashMap::new()));
+        let clients = Arc::new(Mutex::new(HashMap::new()));
+        let feeds = Arc::new(Mutex::new(HashMap::new()));
         Ok(DaemonIdm {
-            receiver,
+            glt,
+            rx_receiver,
+            tx_receiver,
+            tx_sender,
+            tx_raw_sender,
+            snoop_sender,
             task,
-            listeners,
+            clients,
+            feeds,
         })
     }
 
     pub async fn launch(mut self) -> Result<DaemonIdmHandle> {
-        let listeners = self.listeners.clone();
+        let glt = self.glt.clone();
+        let clients = self.clients.clone();
+        let feeds = self.feeds.clone();
+        let tx_sender = self.tx_sender.clone();
+        let snoop_sender = self.snoop_sender.clone();
         let task = tokio::task::spawn(async move {
             let mut buffers: HashMap<u32, BytesMut> = HashMap::new();
-            if let Err(error) = self.process(&mut buffers).await {
+
+            while let Err(error) = self.process(&mut buffers).await {
                 error!("failed to process idm: {}", error);
             }
         });
         Ok(DaemonIdmHandle {
-            listeners,
+            glt,
+            clients,
+            feeds,
+            tx_sender,
+            snoop_sender,
             task: Arc::new(task),
         })
     }
 
     async fn process(&mut self, buffers: &mut HashMap<u32, BytesMut>) -> Result<()> {
         loop {
-            let Some((domid, data)) = self.receiver.recv().await else {
-                break;
-            };
+            select! {
+                x = self.rx_receiver.recv() => match x {
+                    Some((domid, data)) => {
+                        if let Some(data) = data {
+                            let buffer = buffers.entry(domid).or_insert_with_key(|_| BytesMut::new());
+                            buffer.extend_from_slice(&data);
+                            if buffer.len() < 6 {
+                                continue;
+                            }
 
-            let buffer = buffers.entry(domid).or_insert_with_key(|_| BytesMut::new());
-            buffer.extend_from_slice(&data);
-            if buffer.len() < 2 {
-                continue;
-            }
-            let size = (buffer[0] as u16 | (buffer[1] as u16) << 8) as usize;
-            let needed = size + 2;
-            if buffer.len() < needed {
-                continue;
-            }
-            let mut packet = buffer.split_to(needed);
-            packet.advance(2);
-            match IdmPacket::decode(packet) {
-                Ok(packet) => {
-                    let guard = self.listeners.lock().await;
-                    if let Some(sender) = guard.get(&domid) {
-                        if let Err(error) = sender.try_send((domid, packet)) {
-                            warn!("dropped idm packet from domain {}: {}", domid, error);
+                            if buffer[0] != 0xff || buffer[1] != 0xff {
+                                buffer.clear();
+                                continue;
+                            }
+
+                            let size = (buffer[2] as u32 | (buffer[3] as u32) << 8 | (buffer[4] as u32) << 16 | (buffer[5] as u32) << 24) as usize;
+                            let needed = size + 6;
+                            if buffer.len() < needed {
+                                continue;
+                            }
+                            let mut packet = buffer.split_to(needed);
+                            packet.advance(6);
+                            match IdmTransportPacket::decode(packet) {
+                                Ok(packet) => {
+                                    let _ = client_or_create(domid, &self.tx_sender, &self.clients, &self.feeds).await?;
+                                    let guard = self.feeds.lock().await;
+                                    if let Some(feed) = guard.get(&domid) {
+                                        let _ = feed.try_send(packet.clone());
+                                    }
+                                    let _ = self.snoop_sender.send(DaemonIdmSnoopPacket { from: domid, to: 0, packet });
+                                }
+
+                                Err(packet) => {
+                                    warn!("received invalid packet from domain {}: {}", domid, packet);
+                                }
+                            }
+                        } else {
+                            let mut clients = self.clients.lock().await;
+                            let mut feeds = self.feeds.lock().await;
+                            clients.remove(&domid);
+                            feeds.remove(&domid);
                         }
+                    },
+
+                    None => {
+                        break;
+                    }
+                },
+                x = self.tx_receiver.recv() => match x {
+                    Some((domid, packet)) => {
+                        let data = packet.encode_to_vec();
+                        let mut buffer = vec![0u8; 6];
+                        let length = data.len() as u32;
+                        buffer[0] = 0xff;
+                        buffer[1] = 0xff;
+                        buffer[2] = length as u8;
+                        buffer[3] = (length << 8) as u8;
+                        buffer[4] = (length << 16) as u8;
+                        buffer[5] = (length << 24) as u8;
+                        buffer.extend_from_slice(&data);
+                        self.tx_raw_sender.send((domid, buffer)).await?;
+                        let _ = self.snoop_sender.send(DaemonIdmSnoopPacket { from: 0, to: domid, packet });
+                    },
+
+                    None => {
+                        break;
                     }
                 }
-
-                Err(packet) => {
-                    warn!("received invalid packet from domain {}: {}", domid, packet);
-                }
-            }
+            };
         }
         Ok(())
     }
@@ -133,3 +209,54 @@ impl Drop for DaemonIdm {
         self.task.abort();
     }
 }
+
+async fn client_or_create(
+    domid: u32,
+    tx_sender: &Sender<(u32, IdmTransportPacket)>,
+    clients: &ClientMap,
+    feeds: &BackendFeedMap,
+) -> Result<IdmInternalClient> {
+    let mut clients = clients.lock().await;
+    let mut feeds = feeds.lock().await;
+    match clients.entry(domid) {
+        Entry::Occupied(entry) => Ok(entry.get().clone()),
+        Entry::Vacant(entry) => {
+            let (rx_sender, rx_receiver) = channel(100);
+            feeds.insert(domid, rx_sender);
+            let backend = IdmDaemonBackend {
+                domid,
+                rx_receiver,
+                tx_sender: tx_sender.clone(),
+            };
+            let client = IdmInternalClient::new(
+                INTERNAL_IDM_CHANNEL,
+                Box::new(backend) as Box<dyn IdmBackend>,
+            )
+            .await?;
+            entry.insert(client.clone());
+            Ok(client)
+        }
+    }
+}
+
+pub struct IdmDaemonBackend {
+    domid: u32,
+    rx_receiver: Receiver<IdmTransportPacket>,
+    tx_sender: Sender<(u32, IdmTransportPacket)>,
+}
+
+#[async_trait::async_trait]
+impl IdmBackend for IdmDaemonBackend {
+    async fn recv(&mut self) -> Result<IdmTransportPacket> {
+        if let Some(packet) = self.rx_receiver.recv().await {
+            Ok(packet)
+        } else {
+            Err(anyhow!("idm receive channel closed"))
+        }
+    }
+
+    async fn send(&mut self, packet: IdmTransportPacket) -> Result<()> {
+        self.tx_sender.send((self.domid, packet)).await?;
+        Ok(())
+    }
+}

crates/daemon/src/lib.rs

@@ -1,15 +1,19 @@
 use std::{net::SocketAddr, path::PathBuf, str::FromStr};
 
-use anyhow::Result;
-use control::RuntimeControlService;
+use anyhow::{anyhow, Result};
+use console::{DaemonConsole, DaemonConsoleHandle};
+use control::DaemonControlService;
 use db::GuestStore;
 use event::{DaemonEventContext, DaemonEventGenerator};
+use glt::GuestLookupTable;
 use idm::{DaemonIdm, DaemonIdmHandle};
 use krata::{dial::ControlDialAddress, v1::control::control_service_server::ControlServiceServer};
+use krataoci::{packer::service::OciPackerService, registry::OciPlatform};
 use kratart::Runtime;
 use log::info;
 use reconcile::guest::GuestReconciler;
 use tokio::{
+    fs,
     net::UnixListener,
     sync::mpsc::{channel, Sender},
     task::JoinHandle,
@@ -18,60 +22,113 @@ use tokio_stream::wrappers::UnixListenerStream;
 use tonic::transport::{Identity, Server, ServerTlsConfig};
 use uuid::Uuid;
 
+pub mod command;
+pub mod console;
 pub mod control;
 pub mod db;
 pub mod event;
+pub mod glt;
 pub mod idm;
+pub mod metrics;
+pub mod oci;
 pub mod reconcile;
 
 pub struct Daemon {
     store: String,
-    runtime: Runtime,
+    glt: GuestLookupTable,
     guests: GuestStore,
     events: DaemonEventContext,
     guest_reconciler_task: JoinHandle<()>,
     guest_reconciler_notify: Sender<Uuid>,
     generator_task: JoinHandle<()>,
-    _idm: DaemonIdmHandle,
+    idm: DaemonIdmHandle,
+    console: DaemonConsoleHandle,
+    packer: OciPackerService,
 }
 
 const GUEST_RECONCILER_QUEUE_LEN: usize = 1000;
 
 impl Daemon {
-    pub async fn new(store: String, runtime: Runtime) -> Result<Self> {
+    pub async fn new(store: String) -> Result<Self> {
+        let mut image_cache_dir = PathBuf::from(store.clone());
+        image_cache_dir.push("cache");
+        image_cache_dir.push("image");
+        fs::create_dir_all(&image_cache_dir).await?;
+
+        let mut host_uuid_path = PathBuf::from(store.clone());
+        host_uuid_path.push("host.uuid");
+        let host_uuid = if host_uuid_path.is_file() {
+            let content = fs::read_to_string(&host_uuid_path).await?;
+            Uuid::from_str(content.trim()).ok()
+        } else {
+            None
+        };
+
+        let host_uuid = if let Some(host_uuid) = host_uuid {
+            host_uuid
+        } else {
+            let generated = Uuid::new_v4();
+            let mut string = generated.to_string();
+            string.push('\n');
+            fs::write(&host_uuid_path, string).await?;
+            generated
+        };
+
+        let initrd_path = detect_guest_file(&store, "initrd")?;
+        let kernel_path = detect_guest_file(&store, "kernel")?;
+
+        let packer = OciPackerService::new(None, &image_cache_dir, OciPlatform::current()).await?;
+        let runtime = Runtime::new().await?;
+        let glt = GuestLookupTable::new(0, host_uuid);
         let guests_db_path = format!("{}/guests.db", store);
         let guests = GuestStore::open(&PathBuf::from(guests_db_path))?;
         let (guest_reconciler_notify, guest_reconciler_receiver) =
             channel::<Uuid>(GUEST_RECONCILER_QUEUE_LEN);
-        let idm = DaemonIdm::new().await?;
+        let idm = DaemonIdm::new(glt.clone()).await?;
         let idm = idm.launch().await?;
+        let console = DaemonConsole::new(glt.clone()).await?;
+        let console = console.launch().await?;
         let (events, generator) =
             DaemonEventGenerator::new(guests.clone(), guest_reconciler_notify.clone(), idm.clone())
                 .await?;
         let runtime_for_reconciler = runtime.dupe().await?;
-        let guest_reconciler =
-            GuestReconciler::new(guests.clone(), events.clone(), runtime_for_reconciler)?;
+        let guest_reconciler = GuestReconciler::new(
+            glt.clone(),
+            guests.clone(),
+            events.clone(),
+            runtime_for_reconciler,
+            packer.clone(),
+            guest_reconciler_notify.clone(),
+            kernel_path,
+            initrd_path,
+        )?;
 
         let guest_reconciler_task = guest_reconciler.launch(guest_reconciler_receiver).await?;
         let generator_task = generator.launch().await?;
+
         Ok(Self {
             store,
-            runtime,
+            glt,
             guests,
             events,
             guest_reconciler_task,
             guest_reconciler_notify,
             generator_task,
-            _idm: idm,
+            idm,
+            console,
+            packer,
         })
     }
 
     pub async fn listen(&mut self, addr: ControlDialAddress) -> Result<()> {
-        let control_service = RuntimeControlService::new(
+        let control_service = DaemonControlService::new(
+            self.glt.clone(),
             self.events.clone(),
-            self.runtime.clone(),
+            self.console.clone(),
+            self.idm.clone(),
             self.guests.clone(),
             self.guest_reconciler_notify.clone(),
+            self.packer.clone(),
         );
 
         let mut server = Server::builder();
@@ -97,7 +154,7 @@ impl Daemon {
             ControlDialAddress::UnixSocket { path } => {
                 let path = PathBuf::from(path);
                 if path.exists() {
-                    tokio::fs::remove_file(&path).await?;
+                    fs::remove_file(&path).await?;
                 }
                 let listener = UnixListener::bind(path)?;
                 let stream = UnixListenerStream::new(listener);
@@ -128,3 +185,16 @@ impl Drop for Daemon {
         self.generator_task.abort();
     }
 }
+
+fn detect_guest_file(store: &str, name: &str) -> Result<PathBuf> {
+    let mut path = PathBuf::from(format!("{}/guest/{}", store, name));
+    if path.is_file() {
+        return Ok(path);
+    }
+
+    path = PathBuf::from(format!("/usr/share/krata/guest/{}", name));
+    if path.is_file() {
+        return Ok(path);
+    }
+    Err(anyhow!("unable to find required guest file: {}", name))
+}

crates/daemon/src/metrics.rs

@@ -0,0 +1,27 @@
+use krata::{
+    idm::internal::{MetricFormat, MetricNode},
+    v1::common::{GuestMetricFormat, GuestMetricNode},
+};
+
+fn idm_metric_format_to_api(format: MetricFormat) -> GuestMetricFormat {
+    match format {
+        MetricFormat::Unknown => GuestMetricFormat::Unknown,
+        MetricFormat::Bytes => GuestMetricFormat::Bytes,
+        MetricFormat::Integer => GuestMetricFormat::Integer,
+        MetricFormat::DurationSeconds => GuestMetricFormat::DurationSeconds,
+    }
+}
+
+pub fn idm_metric_to_api(node: MetricNode) -> GuestMetricNode {
+    let format = node.format();
+    GuestMetricNode {
+        name: node.name,
+        value: node.value,
+        format: idm_metric_format_to_api(format).into(),
+        children: node
+            .children
+            .into_iter()
+            .map(idm_metric_to_api)
+            .collect::<Vec<_>>(),
+    }
+}

crates/daemon/src/oci.rs

@@ -0,0 +1,79 @@
+use krata::v1::control::{
+    image_progress_indication::Indication, ImageProgress, ImageProgressIndication,
+    ImageProgressIndicationBar, ImageProgressIndicationCompleted, ImageProgressIndicationHidden,
+    ImageProgressIndicationSpinner, ImageProgressLayer, ImageProgressLayerPhase,
+    ImageProgressPhase,
+};
+use krataoci::progress::{
+    OciProgress, OciProgressIndication, OciProgressLayer, OciProgressLayerPhase, OciProgressPhase,
+};
+
+fn convert_oci_progress_indication(indication: OciProgressIndication) -> ImageProgressIndication {
+    ImageProgressIndication {
+        indication: Some(match indication {
+            OciProgressIndication::Hidden => Indication::Hidden(ImageProgressIndicationHidden {}),
+            OciProgressIndication::ProgressBar {
+                message,
+                current,
+                total,
+                bytes,
+            } => Indication::Bar(ImageProgressIndicationBar {
+                message: message.unwrap_or_default(),
+                current,
+                total,
+                is_bytes: bytes,
+            }),
+            OciProgressIndication::Spinner { message } => {
+                Indication::Spinner(ImageProgressIndicationSpinner {
+                    message: message.unwrap_or_default(),
+                })
+            }
+            OciProgressIndication::Completed {
+                message,
+                total,
+                bytes,
+            } => Indication::Completed(ImageProgressIndicationCompleted {
+                message: message.unwrap_or_default(),
+                total: total.unwrap_or(0),
+                is_bytes: bytes,
+            }),
+        }),
+    }
+}
+
+fn convert_oci_layer_progress(layer: OciProgressLayer) -> ImageProgressLayer {
+    ImageProgressLayer {
+        id: layer.id,
+        phase: match layer.phase {
+            OciProgressLayerPhase::Waiting => ImageProgressLayerPhase::Waiting,
+            OciProgressLayerPhase::Downloading => ImageProgressLayerPhase::Downloading,
+            OciProgressLayerPhase::Downloaded => ImageProgressLayerPhase::Downloaded,
+            OciProgressLayerPhase::Extracting => ImageProgressLayerPhase::Extracting,
+            OciProgressLayerPhase::Extracted => ImageProgressLayerPhase::Extracted,
+        }
+        .into(),
+        indication: Some(convert_oci_progress_indication(layer.indication)),
+    }
+}
+
+pub fn convert_oci_progress(oci: OciProgress) -> ImageProgress {
+    ImageProgress {
+        phase: match oci.phase {
+            OciProgressPhase::Started => ImageProgressPhase::Started,
+            OciProgressPhase::Resolving => ImageProgressPhase::Resolving,
+            OciProgressPhase::Resolved => ImageProgressPhase::Resolved,
+            OciProgressPhase::ConfigDownload => ImageProgressPhase::ConfigDownload,
+            OciProgressPhase::LayerDownload => ImageProgressPhase::LayerDownload,
+            OciProgressPhase::Assemble => ImageProgressPhase::Assemble,
+            OciProgressPhase::Pack => ImageProgressPhase::Pack,
+            OciProgressPhase::Complete => ImageProgressPhase::Complete,
+        }
+        .into(),
+        layers: oci
+            .layers
+            .into_values()
+            .map(convert_oci_layer_progress)
+            .collect::<Vec<_>>(),
+        indication: Some(convert_oci_progress_indication(oci.indication)),
+    }
+}

crates/daemon/src/reconcile/guest.rs

@@ -1,246 +0,0 @@
-use std::{collections::HashMap, time::Duration};
-
-use anyhow::{anyhow, Result};
-use krata::v1::{
-    common::{
-        guest_image_spec::Image, Guest, GuestErrorInfo, GuestExitInfo, GuestNetworkState,
-        GuestState, GuestStatus,
-    },
-    control::GuestChangedEvent,
-};
-use kratart::{launch::GuestLaunchRequest, GuestInfo, Runtime};
-use log::{error, info, trace, warn};
-use tokio::{select, sync::mpsc::Receiver, task::JoinHandle, time::sleep};
-use uuid::Uuid;
-
-use crate::{
-    db::GuestStore,
-    event::{DaemonEvent, DaemonEventContext},
-};
-
-pub struct GuestReconciler {
-    guests: GuestStore,
-    events: DaemonEventContext,
-    runtime: Runtime,
-}
-
-impl GuestReconciler {
-    pub fn new(guests: GuestStore, events: DaemonEventContext, runtime: Runtime) -> Result<Self> {
-        Ok(Self {
-            guests,
-            events,
-            runtime,
-        })
-    }
-
-    pub async fn launch(self, mut notify: Receiver<Uuid>) -> Result<JoinHandle<()>> {
-        Ok(tokio::task::spawn(async move {
-            if let Err(error) = self.reconcile_runtime(true).await {
-                error!("runtime reconciler failed: {}", error);
-            }
-
-            loop {
-                select! {
-                    x = notify.recv() => match x {
-                        None => {
-                            break;
-                        },
-
-                        Some(uuid) => {
-                            if let Err(error) = self.reconcile(uuid).await {
-                                error!("failed to reconcile guest {}: {}", uuid, error);
-                            }
-                        }
-                    },
-
-                    _ = sleep(Duration::from_secs(5)) => {
-                        if let Err(error) = self.reconcile_runtime(false).await {
-                            error!("runtime reconciler failed: {}", error);
-                        }
-                    }
-                };
-            }
-        }))
-    }
-
-    pub async fn reconcile_runtime(&self, initial: bool) -> Result<()> {
-        trace!("reconciling runtime");
-        let runtime_guests = self.runtime.list().await?;
-        let stored_guests = self.guests.list().await?;
-        for (uuid, mut stored_guest) in stored_guests {
-            let previous_guest = stored_guest.clone();
-            let runtime_guest = runtime_guests.iter().find(|x| x.uuid == uuid);
-            match runtime_guest {
-                None => {
-                    let mut state = stored_guest.state.as_mut().cloned().unwrap_or_default();
-                    if state.status() == GuestStatus::Started {
-                        state.status = GuestStatus::Starting.into();
-                    }
-                    stored_guest.state = Some(state);
-                }
-
-                Some(runtime) => {
-                    let mut state = stored_guest.state.as_mut().cloned().unwrap_or_default();
-                    if let Some(code) = runtime.state.exit_code {
-                        state.status = GuestStatus::Exited.into();
-                        state.exit_info = Some(GuestExitInfo { code });
-                    } else {
-                        state.status = GuestStatus::Started.into();
-                    }
-                    state.network = Some(guestinfo_to_networkstate(runtime));
-                    stored_guest.state = Some(state);
-                }
-            }
-
-            let changed = stored_guest != previous_guest;
-
-            if changed || initial {
-                self.guests.update(uuid, stored_guest).await?;
-                if let Err(error) = self.reconcile(uuid).await {
-                    error!("failed to reconcile guest {}: {}", uuid, error);
-                }
-            }
-        }
-        Ok(())
-    }
-
-    pub async fn reconcile(&self, uuid: Uuid) -> Result<()> {
-        let Some(mut guest) = self.guests.read(uuid).await? else {
-            warn!(
-                "notified of reconcile for guest {} but it didn't exist",
-                uuid
-            );
-            return Ok(());
-        };
-
-        info!("reconciling guest {}", uuid);
-
-        self.events
-            .send(DaemonEvent::GuestChanged(GuestChangedEvent {
-                guest: Some(guest.clone()),
-            }))?;
-
-        let result = match guest.state.as_ref().map(|x| x.status()).unwrap_or_default() {
-            GuestStatus::Starting => self.start(uuid, &mut guest).await,
-            GuestStatus::Destroying | GuestStatus::Exited => self.destroy(uuid, &mut guest).await,
-            _ => Ok(false),
-        };
-
-        let changed = match result {
-            Ok(changed) => changed,
-            Err(error) => {
-                guest.state = Some(guest.state.as_mut().cloned().unwrap_or_default());
-                guest.state.as_mut().unwrap().status = GuestStatus::Failed.into();
-                guest.state.as_mut().unwrap().error_info = Some(GuestErrorInfo {
-                    message: error.to_string(),
-                });
-                true
-            }
-        };
-
-        info!("reconciled guest {}", uuid);
-
-        let status = guest.state.as_ref().map(|x| x.status()).unwrap_or_default();
-        let destroyed = status == GuestStatus::Destroyed || status == GuestStatus::Failed;
-
-        if changed {
-            let event = DaemonEvent::GuestChanged(GuestChangedEvent {
-                guest: Some(guest.clone()),
-            });
-
-            if destroyed {
-                self.guests.remove(uuid).await?;
-            } else {
-                self.guests.update(uuid, guest.clone()).await?;
-            }
-
-            self.events.send(event)?;
-        }
-
-        Ok(())
-    }
-
-    async fn start(&self, uuid: Uuid, guest: &mut Guest) -> Result<bool> {
-        let Some(ref spec) = guest.spec else {
-            return Err(anyhow!("guest spec not specified"));
-        };
-
-        let Some(ref image) = spec.image else {
-            return Err(anyhow!("image spec not provided"));
-        };
-        let oci = match image.image {
-            Some(Image::Oci(ref oci)) => oci,
-            None => {
-                return Err(anyhow!("oci spec not specified"));
-            }
-        };
-
-        let task = spec.task.as_ref().cloned().unwrap_or_default();
-
-        let info = self
-            .runtime
-            .launch(GuestLaunchRequest {
-                uuid: Some(uuid),
-                name: if spec.name.is_empty() {
-                    None
-                } else {
-                    Some(&spec.name)
-                },
-                image: &oci.image,
-                vcpus: spec.vcpus,
-                mem: spec.mem,
-                env: task
-                    .environment
-                    .iter()
-                    .map(|x| (x.key.clone(), x.value.clone()))
-                    .collect::<HashMap<_, _>>(),
-                run: empty_vec_optional(task.command.clone()),
-                debug: false,
-            })
-            .await?;
-        info!("started guest {}", uuid);
-        guest.state = Some(GuestState {
-            status: GuestStatus::Started.into(),
-            network: Some(guestinfo_to_networkstate(&info)),
-            exit_info: None,
-            error_info: None,
-            domid: info.domid,
-        });
-        Ok(true)
-    }
-
-    async fn destroy(&self, uuid: Uuid, guest: &mut Guest) -> Result<bool> {
-        if let Err(error) = self.runtime.destroy(uuid).await {
-            trace!("failed to destroy runtime guest {}: {}", uuid, error);
-        }
-
-        info!("destroyed guest {}", uuid);
-        guest.state = Some(GuestState {
-            status: GuestStatus::Destroyed.into(),
-            network: None,
-            exit_info: None,
-            error_info: None,
-            domid: guest.state.as_ref().map(|x| x.domid).unwrap_or(u32::MAX),
-        });
-        Ok(true)
-    }
-}
-
-fn empty_vec_optional<T>(value: Vec<T>) -> Option<Vec<T>> {
-    if value.is_empty() {
-        None
-    } else {
-        Some(value)
-    }
-}
-
-fn guestinfo_to_networkstate(info: &GuestInfo) -> GuestNetworkState {
-    GuestNetworkState {
-        guest_ipv4: info.guest_ipv4.map(|x| x.to_string()).unwrap_or_default(),
-        guest_ipv6: info.guest_ipv6.map(|x| x.to_string()).unwrap_or_default(),
-        guest_mac: info.guest_mac.as_ref().cloned().unwrap_or_default(),
-        gateway_ipv4: info.gateway_ipv4.map(|x| x.to_string()).unwrap_or_default(),
-        gateway_ipv6: info.gateway_ipv6.map(|x| x.to_string()).unwrap_or_default(),
-        gateway_mac: info.gateway_mac.as_ref().cloned().unwrap_or_default(),
-    }
-}

crates/daemon/src/reconcile/guest/mod.rs

@@ -0,0 +1,348 @@
+use std::{
+    collections::{hash_map::Entry, HashMap},
+    path::PathBuf,
+    sync::Arc,
+    time::Duration,
+};
+
+use anyhow::Result;
+use krata::v1::{
+    common::{Guest, GuestErrorInfo, GuestExitInfo, GuestNetworkState, GuestState, GuestStatus},
+    control::GuestChangedEvent,
+};
+use krataoci::packer::service::OciPackerService;
+use kratart::{GuestInfo, Runtime};
+use log::{error, info, trace, warn};
+use tokio::{
+    select,
+    sync::{
+        mpsc::{channel, Receiver, Sender},
+        Mutex, RwLock,
+    },
+    task::JoinHandle,
+    time::sleep,
+};
+use uuid::Uuid;
+
+use crate::{
+    db::GuestStore,
+    event::{DaemonEvent, DaemonEventContext},
+    glt::GuestLookupTable,
+};
+
+use self::start::GuestStarter;
+
+mod start;
+
+const PARALLEL_LIMIT: u32 = 5;
+
+#[derive(Debug)]
+enum GuestReconcilerResult {
+    Unchanged,
+    Changed { rerun: bool },
+}
+
+struct GuestReconcilerEntry {
+    task: JoinHandle<()>,
+    sender: Sender<()>,
+}
+
+impl Drop for GuestReconcilerEntry {
+    fn drop(&mut self) {
+        self.task.abort();
+    }
+}
+
+#[derive(Clone)]
+pub struct GuestReconciler {
+    glt: GuestLookupTable,
+    guests: GuestStore,
+    events: DaemonEventContext,
+    runtime: Runtime,
+    packer: OciPackerService,
+    kernel_path: PathBuf,
+    initrd_path: PathBuf,
+    tasks: Arc<Mutex<HashMap<Uuid, GuestReconcilerEntry>>>,
+    guest_reconciler_notify: Sender<Uuid>,
+    reconcile_lock: Arc<RwLock<()>>,
+}
+
+impl GuestReconciler {
+    #[allow(clippy::too_many_arguments)]
+    pub fn new(
+        glt: GuestLookupTable,
+        guests: GuestStore,
+        events: DaemonEventContext,
+        runtime: Runtime,
+        packer: OciPackerService,
+        guest_reconciler_notify: Sender<Uuid>,
+        kernel_path: PathBuf,
+        initrd_path: PathBuf,
+    ) -> Result<Self> {
+        Ok(Self {
+            glt,
+            guests,
+            events,
+            runtime,
+            packer,
+            kernel_path,
+            initrd_path,
+            tasks: Arc::new(Mutex::new(HashMap::new())),
+            guest_reconciler_notify,
+            reconcile_lock: Arc::new(RwLock::with_max_readers((), PARALLEL_LIMIT)),
+        })
+    }
+
+    pub async fn launch(self, mut notify: Receiver<Uuid>) -> Result<JoinHandle<()>> {
+        Ok(tokio::task::spawn(async move {
+            if let Err(error) = self.reconcile_runtime(true).await {
+                error!("runtime reconciler failed: {}", error);
+            }
+
+            loop {
+                select! {
+                    x = notify.recv() => match x {
+                        None => {
+                            break;
+                        },
+
+                        Some(uuid) => {
+                            if let Err(error) = self.launch_task_if_needed(uuid).await {
+                                error!("failed to start guest reconciler task {}: {}", uuid, error);
+                            }
+
+                            let map = self.tasks.lock().await;
+                            if let Some(entry) = map.get(&uuid) {
+                                if let Err(error) = entry.sender.send(()).await {
+                                    error!("failed to notify guest reconciler task {}: {}", uuid, error);
+                                }
+                            }
+                        }
+                    },
+
+                    _ = sleep(Duration::from_secs(5)) => {
+                        if let Err(error) = self.reconcile_runtime(false).await {
+                            error!("runtime reconciler failed: {}", error);
+                        }
+                    }
+                };
+            }
+        }))
+    }
+
+    pub async fn reconcile_runtime(&self, initial: bool) -> Result<()> {
+        let _permit = self.reconcile_lock.write().await;
+        trace!("reconciling runtime");
+        let runtime_guests = self.runtime.list().await?;
+        let stored_guests = self.guests.list().await?;
+
+        let non_existent_guests = runtime_guests
+            .iter()
+            .filter(|x| !stored_guests.iter().any(|g| *g.0 == x.uuid))
+            .collect::<Vec<_>>();
+
+        for guest in non_existent_guests {
+            warn!("destroying unknown runtime guest {}", guest.uuid);
+            if let Err(error) = self.runtime.destroy(guest.uuid).await {
+                error!(
+                    "failed to destroy unknown runtime guest {}: {}",
+                    guest.uuid, error
+                );
+            }
+            self.guests.remove(guest.uuid).await?;
+        }
+
+        for (uuid, mut stored_guest) in stored_guests {
+            let previous_guest = stored_guest.clone();
+            let runtime_guest = runtime_guests.iter().find(|x| x.uuid == uuid);
+            match runtime_guest {
+                None => {
+                    let mut state = stored_guest.state.as_mut().cloned().unwrap_or_default();
+                    if state.status() == GuestStatus::Started {
+                        state.status = GuestStatus::Starting.into();
+                    }
+                    stored_guest.state = Some(state);
+                }
+
+                Some(runtime) => {
+                    self.glt.associate(uuid, runtime.domid).await;
+                    let mut state = stored_guest.state.as_mut().cloned().unwrap_or_default();
+                    if let Some(code) = runtime.state.exit_code {
+                        state.status = GuestStatus::Exited.into();
+                        state.exit_info = Some(GuestExitInfo { code });
+                    } else {
+                        state.status = GuestStatus::Started.into();
+                    }
+                    state.network = Some(guestinfo_to_networkstate(runtime));
+                    stored_guest.state = Some(state);
+                }
+            }
+
+            let changed = stored_guest != previous_guest;
+
+            if changed || initial {
+                self.guests.update(uuid, stored_guest).await?;
+                let _ = self.guest_reconciler_notify.try_send(uuid);
+            }
+        }
+        Ok(())
+    }
+
+    pub async fn reconcile(&self, uuid: Uuid) -> Result<bool> {
+        let _runtime_reconcile_permit = self.reconcile_lock.read().await;
+        let Some(mut guest) = self.guests.read(uuid).await? else {
+            warn!(
+                "notified of reconcile for guest {} but it didn't exist",
+                uuid
+            );
+            return Ok(false);
+        };
+
+        info!("reconciling guest {}", uuid);
+
+        self.events
+            .send(DaemonEvent::GuestChanged(GuestChangedEvent {
+                guest: Some(guest.clone()),
+            }))?;
+
+        let start_status = guest.state.as_ref().map(|x| x.status()).unwrap_or_default();
+        let result = match start_status {
+            GuestStatus::Starting => self.start(uuid, &mut guest).await,
+            GuestStatus::Exited => self.exited(&mut guest).await,
+            GuestStatus::Destroying => self.destroy(uuid, &mut guest).await,
+            _ => Ok(GuestReconcilerResult::Unchanged),
+        };
+
+        let result = match result {
+            Ok(result) => result,
+            Err(error) => {
+                guest.state = Some(guest.state.as_mut().cloned().unwrap_or_default());
+                guest.state.as_mut().unwrap().status = GuestStatus::Failed.into();
+                guest.state.as_mut().unwrap().error_info = Some(GuestErrorInfo {
+                    message: error.to_string(),
+                });
+                warn!("failed to start guest {}: {}", guest.id, error);
+                GuestReconcilerResult::Changed { rerun: false }
+            }
+        };
+
+        info!("reconciled guest {}", uuid);
+
+        let status = guest.state.as_ref().map(|x| x.status()).unwrap_or_default();
+        let destroyed = status == GuestStatus::Destroyed;
+
+        let rerun = if let GuestReconcilerResult::Changed { rerun } = result {
+            let event = DaemonEvent::GuestChanged(GuestChangedEvent {
+                guest: Some(guest.clone()),
+            });
+
+            if destroyed {
+                self.guests.remove(uuid).await?;
+                let mut map = self.tasks.lock().await;
+                map.remove(&uuid);
+            } else {
+                self.guests.update(uuid, guest.clone()).await?;
+            }
+
+            self.events.send(event)?;
+            rerun
+        } else {
+            false
+        };
+
+        Ok(rerun)
+    }
+
+    async fn start(&self, uuid: Uuid, guest: &mut Guest) -> Result<GuestReconcilerResult> {
+        let starter = GuestStarter {
+            kernel_path: &self.kernel_path,
+            initrd_path: &self.initrd_path,
+            packer: &self.packer,
+            glt: &self.glt,
+            runtime: &self.runtime,
+        };
+        starter.start(uuid, guest).await
+    }
+
+    async fn exited(&self, guest: &mut Guest) -> Result<GuestReconcilerResult> {
+        if let Some(ref mut state) = guest.state {
+            state.set_status(GuestStatus::Destroying);
+            Ok(GuestReconcilerResult::Changed { rerun: true })
+        } else {
+            Ok(GuestReconcilerResult::Unchanged)
+        }
+    }
+
+    async fn destroy(&self, uuid: Uuid, guest: &mut Guest) -> Result<GuestReconcilerResult> {
+        if let Err(error) = self.runtime.destroy(uuid).await {
+            trace!("failed to destroy runtime guest {}: {}", uuid, error);
+        }
+
+        let domid = guest.state.as_ref().map(|x| x.domid);
+
+        if let Some(domid) = domid {
+            self.glt.remove(uuid, domid).await;
+        }
+
+        info!("destroyed guest {}", uuid);
+        guest.state = Some(GuestState {
+            status: GuestStatus::Destroyed.into(),
+            network: None,
+            exit_info: None,
+            error_info: None,
+            host: self.glt.host_uuid().to_string(),
+            domid: domid.unwrap_or(u32::MAX),
+        });
+        Ok(GuestReconcilerResult::Changed { rerun: false })
+    }
+
+    async fn launch_task_if_needed(&self, uuid: Uuid) -> Result<()> {
+        let mut map = self.tasks.lock().await;
+        match map.entry(uuid) {
+            Entry::Occupied(_) => {}
+            Entry::Vacant(entry) => {
+                entry.insert(self.launch_task(uuid).await?);
+            }
+        }
+        Ok(())
+    }
+
+    async fn launch_task(&self, uuid: Uuid) -> Result<GuestReconcilerEntry> {
+        let this = self.clone();
+        let (sender, mut receiver) = channel(10);
+        let task = tokio::task::spawn(async move {
+            'notify_loop: loop {
+                if receiver.recv().await.is_none() {
+                    break 'notify_loop;
+                }
+
+                'rerun_loop: loop {
+                    let rerun = match this.reconcile(uuid).await {
+                        Ok(rerun) => rerun,
+                        Err(error) => {
+                            error!("failed to reconcile guest {}: {}", uuid, error);
+                            false
+                        }
+                    };
+
+                    if rerun {
+                        continue 'rerun_loop;
+                    }
+                    break 'rerun_loop;
+                }
+            }
+        });
+        Ok(GuestReconcilerEntry { task, sender })
+    }
+}
+
+pub fn guestinfo_to_networkstate(info: &GuestInfo) -> GuestNetworkState {
+    GuestNetworkState {
+        guest_ipv4: info.guest_ipv4.map(|x| x.to_string()).unwrap_or_default(),
+        guest_ipv6: info.guest_ipv6.map(|x| x.to_string()).unwrap_or_default(),
+        guest_mac: info.guest_mac.as_ref().cloned().unwrap_or_default(),
+        gateway_ipv4: info.gateway_ipv4.map(|x| x.to_string()).unwrap_or_default(),
+        gateway_ipv6: info.gateway_ipv6.map(|x| x.to_string()).unwrap_or_default(),
+        gateway_mac: info.gateway_mac.as_ref().cloned().unwrap_or_default(),
+    }
+}

crates/daemon/src/reconcile/guest/start.rs

@@ -0,0 +1,182 @@
+use std::collections::HashMap;
+use std::path::{Path, PathBuf};
+
+use anyhow::{anyhow, Result};
+use futures::StreamExt;
+use krata::launchcfg::LaunchPackedFormat;
+use krata::v1::common::GuestOciImageSpec;
+use krata::v1::common::{guest_image_spec::Image, Guest, GuestState, GuestStatus, OciImageFormat};
+use krataoci::packer::{service::OciPackerService, OciPackedFormat};
+use kratart::{launch::GuestLaunchRequest, Runtime};
+use log::info;
+
+use tokio::fs::{self, File};
+use tokio::io::AsyncReadExt;
+use tokio_tar::Archive;
+use uuid::Uuid;
+
+use crate::{
+    glt::GuestLookupTable,
+    reconcile::guest::{guestinfo_to_networkstate, GuestReconcilerResult},
+};
+
+// if a kernel is >= 100MB, that's kinda scary.
+const OCI_SPEC_TAR_FILE_MAX_SIZE: usize = 100 * 1024 * 1024;
+
+pub struct GuestStarter<'a> {
+    pub kernel_path: &'a Path,
+    pub initrd_path: &'a Path,
+    pub packer: &'a OciPackerService,
+    pub glt: &'a GuestLookupTable,
+    pub runtime: &'a Runtime,
+}
+
+impl GuestStarter<'_> {
+    pub async fn oci_spec_tar_read_file(
+        &self,
+        file: &Path,
+        oci: &GuestOciImageSpec,
+    ) -> Result<Vec<u8>> {
+        if oci.format() != OciImageFormat::Tar {
+            return Err(anyhow!(
+                "oci image spec for {} is required to be in tar format",
+                oci.digest
+            ));
+        }
+
+        let image = self
+            .packer
+            .recall(&oci.digest, OciPackedFormat::Tar)
+            .await?;
+
+        let Some(image) = image else {
+            return Err(anyhow!("image {} was not found in tar format", oci.digest));
+        };
+
+        let mut archive = Archive::new(File::open(&image.path).await?);
+        let mut entries = archive.entries()?;
+        while let Some(entry) = entries.next().await {
+            let mut entry = entry?;
+            let path = entry.path()?;
+            if entry.header().size()? as usize > OCI_SPEC_TAR_FILE_MAX_SIZE {
+                return Err(anyhow!(
+                    "file {} in image {} is larger than the size limit",
+                    file.to_string_lossy(),
+                    oci.digest
+                ));
+            }
+            if path == file {
+                let mut buffer = Vec::new();
+                entry.read_to_end(&mut buffer).await?;
+                return Ok(buffer);
+            }
+        }
+        Err(anyhow!(
+            "unable to find file {} in image {}",
+            file.to_string_lossy(),
+            oci.digest
+        ))
+    }
+
+    pub async fn start(&self, uuid: Uuid, guest: &mut Guest) -> Result<GuestReconcilerResult> {
+        let Some(ref spec) = guest.spec else {
+            return Err(anyhow!("guest spec not specified"));
+        };
+
+        let Some(ref image) = spec.image else {
+            return Err(anyhow!("image spec not provided"));
+        };
+        let oci = match image.image {
+            Some(Image::Oci(ref oci)) => oci,
+            None => {
+                return Err(anyhow!("oci spec not specified"));
+            }
+        };
+        let task = spec.task.as_ref().cloned().unwrap_or_default();
+
+        let image = self
+            .packer
+            .recall(
+                &oci.digest,
+                match oci.format() {
+                    OciImageFormat::Unknown => OciPackedFormat::Squashfs,
+                    OciImageFormat::Squashfs => OciPackedFormat::Squashfs,
+                    OciImageFormat::Erofs => OciPackedFormat::Erofs,
+                    OciImageFormat::Tar => {
+                        return Err(anyhow!("tar image format is not supported for guests"));
+                    }
+                },
+            )
+            .await?;
+
+        let Some(image) = image else {
+            return Err(anyhow!(
+                "image {} in the requested format did not exist",
+                oci.digest
+            ));
+        };
+
+        let kernel = if let Some(ref spec) = spec.kernel {
+            let Some(Image::Oci(ref oci)) = spec.image else {
+                return Err(anyhow!("kernel image spec must be an oci image"));
+            };
+            self.oci_spec_tar_read_file(&PathBuf::from("kernel/image"), oci)
+                .await?
+        } else {
+            fs::read(&self.kernel_path).await?
+        };
+        let initrd = if let Some(ref spec) = spec.initrd {
+            let Some(Image::Oci(ref oci)) = spec.image else {
+                return Err(anyhow!("initrd image spec must be an oci image"));
+            };
+            self.oci_spec_tar_read_file(&PathBuf::from("krata/initrd"), oci)
+                .await?
+        } else {
+            fs::read(&self.initrd_path).await?
+        };
+
+        let info = self
+            .runtime
+            .launch(GuestLaunchRequest {
+                format: LaunchPackedFormat::Squashfs,
+                uuid: Some(uuid),
+                name: if spec.name.is_empty() {
+                    None
+                } else {
+                    Some(spec.name.clone())
+                },
+                image,
+                kernel,
+                initrd,
+                vcpus: spec.vcpus,
+                mem: spec.mem,
+                env: task
+                    .environment
+                    .iter()
+                    .map(|x| (x.key.clone(), x.value.clone()))
+                    .collect::<HashMap<_, _>>(),
+                run: empty_vec_optional(task.command.clone()),
+                debug: false,
+            })
+            .await?;
+        self.glt.associate(uuid, info.domid).await;
+        info!("started guest {}", uuid);
+        guest.state = Some(GuestState {
+            status: GuestStatus::Started.into(),
+            network: Some(guestinfo_to_networkstate(&info)),
+            exit_info: None,
+            error_info: None,
+            host: self.glt.host_uuid().to_string(),
+            domid: info.domid,
+        });
+        Ok(GuestReconcilerResult::Changed { rerun: false })
+    }
+}
+
+fn empty_vec_optional<T>(value: Vec<T>) -> Option<Vec<T>> {
+    if value.is_empty() {
+        None
+    } else {
+        Some(value)
+    }
+}

crates/guest/Cargo.toml

@@ -2,7 +2,7 @@
 name = "krata-guest"
 description = "Guest services for the krata hypervisor."
 license.workspace = true
-version= "0.0.3"
+version.workspace = true
 homepage.workspace = true
 repository.workspace = true
 edition = "2021"
@@ -10,11 +10,12 @@ resolver = "2"
 
 [dependencies]
 anyhow = { workspace = true }
+cgroups-rs = { workspace = true }
 env_logger = { workspace = true }
 futures = { workspace = true }
 ipnetwork = { workspace = true }
-krata = { path = "../krata", version = "^0.0.3" }
-krata-xenstore = { path = "../xen/xenstore", version = "^0.0.3" }
+krata = { path = "../krata", version = "^0.0.10" }
+krata-xenstore = { path = "../xen/xenstore", version = "^0.0.10" }
 libc = { workspace = true }
 log = { workspace = true }
 nix = { workspace = true, features = ["ioctl", "process", "fs"] }
@@ -24,8 +25,8 @@ rtnetlink = { workspace = true }
 serde = { workspace = true }
 serde_json = { workspace = true }
 sys-mount = { workspace = true }
+sysinfo = { workspace = true }
 tokio = { workspace = true }
-walkdir = { workspace = true }
 
 [lib]
 name = "krataguest"

crates/guest/bin/init.rs

@@ -23,6 +23,8 @@ async fn main() -> Result<()> {
     if let Err(error) = guest.init().await {
         error!("failed to initialize guest: {}", error);
         death(127).await?;
+        return Ok(());
     }
+    death(1).await?;
     Ok(())
 }

crates/guest/src/background.rs

@@ -1,42 +1,91 @@
 use crate::{
     childwait::{ChildEvent, ChildWait},
     death,
+    exec::GuestExecTask,
+    metrics::MetricsCollector,
 };
 use anyhow::Result;
+use cgroups_rs::Cgroup;
 use krata::idm::{
-    client::IdmClient,
-    protocol::{idm_event::Event, IdmEvent, IdmExitEvent, IdmPacket},
+    client::{IdmClientStreamResponseHandle, IdmInternalClient},
+    internal::{
+        event::Event as EventType, request::Request as RequestType,
+        response::Response as ResponseType, Event, ExecStreamResponseUpdate, ExitEvent,
+        MetricsResponse, PingResponse, Request, Response,
+    },
 };
-use log::error;
+use log::debug;
 use nix::unistd::Pid;
-use tokio::select;
+use tokio::{select, sync::broadcast};
 
 pub struct GuestBackground {
-    idm: IdmClient,
+    idm: IdmInternalClient,
     child: Pid,
+    _cgroup: Cgroup,
     wait: ChildWait,
 }
 
 impl GuestBackground {
-    pub async fn new(idm: IdmClient, child: Pid) -> Result<GuestBackground> {
+    pub async fn new(
+        idm: IdmInternalClient,
+        cgroup: Cgroup,
+        child: Pid,
+    ) -> Result<GuestBackground> {
         Ok(GuestBackground {
             idm,
             child,
+            _cgroup: cgroup,
             wait: ChildWait::new()?,
         })
     }
 
     pub async fn run(&mut self) -> Result<()> {
+        let mut event_subscription = self.idm.subscribe().await?;
+        let mut requests_subscription = self.idm.requests().await?;
+        let mut request_streams_subscription = self.idm.request_streams().await?;
         loop {
             select! {
-                x = self.idm.receiver.recv() => match x {
-                    Some(_packet) => {
-
+                x = event_subscription.recv() => match x {
+                    Ok(_event) => {
                     },
 
-                    None => {
-                        error!("idm packet channel closed");
+                    Err(broadcast::error::RecvError::Closed) => {
+                        debug!("idm packet channel closed");
                         break;
+                    },
+
+                    _ => {
+                        continue;
+                    }
+                },
+
+                x = requests_subscription.recv() => match x {
+                    Ok((id, request)) => {
+                        self.handle_idm_request(id, request).await?;
+                    },
+
+                    Err(broadcast::error::RecvError::Closed) => {
+                        debug!("idm packet channel closed");
+                        break;
+                    },
+
+                    _ => {
+                        continue;
+                    }
+                },
+
+                x = request_streams_subscription.recv() => match x {
+                    Ok(handle) => {
+                        self.handle_idm_stream_request(handle).await?;
+                    },
+
+                    Err(broadcast::error::RecvError::Closed) => {
+                        debug!("idm packet channel closed");
+                        break;
+                    },
+
+                    _ => {
+                        continue;
                     }
                 },
 
@@ -51,14 +100,65 @@ impl GuestBackground {
         Ok(())
     }
 
+    async fn handle_idm_request(&mut self, id: u64, packet: Request) -> Result<()> {
+        match packet.request {
+            Some(RequestType::Ping(_)) => {
+                self.idm
+                    .respond(
+                        id,
+                        Response {
+                            response: Some(ResponseType::Ping(PingResponse {})),
+                        },
+                    )
+                    .await?;
+            }
+
+            Some(RequestType::Metrics(_)) => {
+                let metrics = MetricsCollector::new()?;
+                let root = metrics.collect()?;
+                let response = Response {
+                    response: Some(ResponseType::Metrics(MetricsResponse { root: Some(root) })),
+                };
+
+                self.idm.respond(id, response).await?;
+            }
+
+            _ => {}
+        }
+        Ok(())
+    }
+
+    async fn handle_idm_stream_request(
+        &mut self,
+        handle: IdmClientStreamResponseHandle<Request>,
+    ) -> Result<()> {
+        if let Some(RequestType::ExecStream(_)) = &handle.initial.request {
+            tokio::task::spawn(async move {
+                let exec = GuestExecTask { handle };
+                if let Err(error) = exec.run().await {
+                    let _ = exec
+                        .handle
+                        .respond(Response {
+                            response: Some(ResponseType::ExecStream(ExecStreamResponseUpdate {
+                                exited: true,
+                                error: error.to_string(),
+                                exit_code: -1,
+                                stdout: vec![],
+                                stderr: vec![],
+                            })),
+                        })
+                        .await;
+                }
+            });
+        }
+        Ok(())
+    }
+
     async fn child_event(&mut self, event: ChildEvent) -> Result<()> {
         if event.pid == self.child {
             self.idm
-                .sender
-                .send(IdmPacket {
-                    event: Some(IdmEvent {
-                        event: Some(Event::Exit(IdmExitEvent { code: event.status })),
-                    }),
+                .emit(Event {
+                    event: Some(EventType::Exit(ExitEvent { code: event.status })),
                 })
                 .await?;
             death(event.status).await?;

crates/guest/src/exec.rs

@@ -0,0 +1,172 @@
+use std::{collections::HashMap, process::Stdio};
+
+use anyhow::{anyhow, Result};
+use krata::idm::{
+    client::IdmClientStreamResponseHandle,
+    internal::{
+        exec_stream_request_update::Update, request::Request as RequestType,
+        ExecStreamResponseUpdate,
+    },
+    internal::{response::Response as ResponseType, Request, Response},
+};
+use tokio::{
+    io::{AsyncReadExt, AsyncWriteExt},
+    join,
+    process::Command,
+};
+
+pub struct GuestExecTask {
+    pub handle: IdmClientStreamResponseHandle<Request>,
+}
+
+impl GuestExecTask {
+    pub async fn run(&self) -> Result<()> {
+        let mut receiver = self.handle.take().await?;
+
+        let Some(ref request) = self.handle.initial.request else {
+            return Err(anyhow!("request was empty"));
+        };
+
+        let RequestType::ExecStream(update) = request else {
+            return Err(anyhow!("request was not an exec update"));
+        };
+
+        let Some(Update::Start(ref start)) = update.update else {
+            return Err(anyhow!("first request did not contain a start update"));
+        };
+
+        let mut cmd = start.command.clone();
+        if cmd.is_empty() {
+            return Err(anyhow!("command line was empty"));
+        }
+        let exe = cmd.remove(0);
+        let mut env = HashMap::new();
+        for entry in &start.environment {
+            env.insert(entry.key.clone(), entry.value.clone());
+        }
+
+        if !env.contains_key("PATH") {
+            env.insert(
+                "PATH".to_string(),
+                "/bin:/usr/bin:/usr/local/bin".to_string(),
+            );
+        }
+
+        let dir = if start.working_directory.is_empty() {
+            "/".to_string()
+        } else {
+            start.working_directory.clone()
+        };
+
+        let mut child = Command::new(exe)
+            .args(cmd)
+            .envs(env)
+            .current_dir(dir)
+            .stdin(Stdio::piped())
+            .stdout(Stdio::piped())
+            .stderr(Stdio::piped())
+            .kill_on_drop(true)
+            .spawn()
+            .map_err(|error| anyhow!("failed to spawn: {}", error))?;
+
+        let mut stdin = child
+            .stdin
+            .take()
+            .ok_or_else(|| anyhow!("stdin was missing"))?;
+        let mut stdout = child
+            .stdout
+            .take()
+            .ok_or_else(|| anyhow!("stdout was missing"))?;
+        let mut stderr = child
+            .stderr
+            .take()
+            .ok_or_else(|| anyhow!("stderr was missing"))?;
+
+        let stdout_handle = self.handle.clone();
+        let stdout_task = tokio::task::spawn(async move {
+            let mut stdout_buffer = vec![0u8; 8 * 1024];
+            loop {
+                let Ok(size) = stdout.read(&mut stdout_buffer).await else {
+                    break;
+                };
+                if size > 0 {
+                    let response = Response {
+                        response: Some(ResponseType::ExecStream(ExecStreamResponseUpdate {
+                            exited: false,
+                            exit_code: 0,
+                            error: String::new(),
+                            stdout: stdout_buffer[0..size].to_vec(),
+                            stderr: vec![],
+                        })),
+                    };
+                    let _ = stdout_handle.respond(response).await;
+                } else {
+                    break;
+                }
+            }
+        });
+
+        let stderr_handle = self.handle.clone();
+        let stderr_task = tokio::task::spawn(async move {
+            let mut stderr_buffer = vec![0u8; 8 * 1024];
+            loop {
+                let Ok(size) = stderr.read(&mut stderr_buffer).await else {
+                    break;
+                };
+                if size > 0 {
+                    let response = Response {
+                        response: Some(ResponseType::ExecStream(ExecStreamResponseUpdate {
+                            exited: false,
+                            exit_code: 0,
+                            error: String::new(),
+                            stdout: vec![],
+                            stderr: stderr_buffer[0..size].to_vec(),
+                        })),
+                    };
+                    let _ = stderr_handle.respond(response).await;
+                } else {
+                    break;
+                }
+            }
+        });
+
+        let stdin_task = tokio::task::spawn(async move {
+            loop {
+                let Some(request) = receiver.recv().await else {
+                    break;
+                };
+
+                let Some(RequestType::ExecStream(update)) = request.request else {
+                    continue;
+                };
+
+                let Some(Update::Stdin(update)) = update.update else {
+                    continue;
+                };
+
+                if stdin.write_all(&update.data).await.is_err() {
+                    break;
+                }
+            }
+        });
+
+        let exit = child.wait().await?;
+        let code = exit.code().unwrap_or(-1);
+
+        let _ = join!(stdout_task, stderr_task);
+        stdin_task.abort();
+
+        let response = Response {
+            response: Some(ResponseType::ExecStream(ExecStreamResponseUpdate {
+                exited: true,
+                exit_code: code,
+                error: String::new(),
+                stdout: vec![],
+                stderr: vec![],
+            })),
+        };
+        self.handle.respond(response).await?;
+
+        Ok(())
+    }
+}

crates/guest/src/init.rs

@@ -1,10 +1,12 @@
 use anyhow::{anyhow, Result};
+use cgroups_rs::{Cgroup, CgroupPid};
 use futures::stream::TryStreamExt;
 use ipnetwork::IpNetwork;
 use krata::ethtool::EthtoolHandle;
-use krata::idm::client::IdmClient;
-use krata::launchcfg::{LaunchInfo, LaunchNetwork};
-use libc::{setsid, TIOCSCTTY};
+use krata::idm::client::IdmInternalClient;
+use krata::idm::internal::INTERNAL_IDM_CHANNEL;
+use krata::launchcfg::{LaunchInfo, LaunchNetwork, LaunchPackedFormat};
+use libc::{sethostname, setsid, TIOCSCTTY};
 use log::{trace, warn};
 use nix::ioctl_write_int_bad;
 use nix::unistd::{dup2, execve, fork, ForkResult, Pid};
@@ -16,14 +18,12 @@ use std::fs::{File, OpenOptions, Permissions};
 use std::io;
 use std::net::{Ipv4Addr, Ipv6Addr};
 use std::os::fd::AsRawFd;
-use std::os::linux::fs::MetadataExt;
 use std::os::unix::ffi::OsStrExt;
-use std::os::unix::fs::{chroot, symlink, PermissionsExt};
+use std::os::unix::fs::{chroot, PermissionsExt};
 use std::path::{Path, PathBuf};
 use std::str::FromStr;
 use sys_mount::{FilesystemType, Mount, MountFlags};
 use tokio::fs;
-use walkdir::WalkDir;
 
 use crate::background::GuestBackground;
 
@@ -78,21 +78,42 @@ impl GuestInit {
             Err(error) => warn!("failed to open console: {}", error),
         };
 
-        let idm = IdmClient::open("/dev/hvc1")
+        let idm = IdmInternalClient::open(INTERNAL_IDM_CHANNEL, "/dev/hvc1")
             .await
             .map_err(|x| anyhow!("failed to open idm client: {}", x))?;
-        self.mount_squashfs_images().await?;
+        self.mount_config_image().await?;
 
         let config = self.parse_image_config().await?;
         let launch = self.parse_launch_config().await?;
 
+        self.mount_root_image(launch.root.format.clone()).await?;
+
         self.mount_new_root().await?;
-        self.nuke_initrd().await?;
         self.bind_new_root().await?;
 
+        if let Some(hostname) = launch.hostname.clone() {
+            let result = unsafe {
+                sethostname(
+                    hostname.as_bytes().as_ptr() as *mut libc::c_char,
+                    hostname.len(),
+                )
+            };
+            if result != 0 {
+                warn!("failed to set hostname: {}", result);
+            }
+
+            let etc = PathBuf::from_str("/etc")?;
+            if !etc.exists() {
+                fs::create_dir(&etc).await?;
+            }
+            let mut etc_hostname = etc;
+            etc_hostname.push("hostname");
+            fs::write(&etc_hostname, hostname + "\n").await?;
+        }
+
         if let Some(network) = &launch.network {
             trace!("initializing network");
-            if let Err(error) = self.network_setup(network).await {
+            if let Err(error) = self.network_setup(&launch, network).await {
                 warn!("failed to initialize network: {}", error);
             }
         }
@@ -112,14 +133,20 @@ impl GuestInit {
         trace!("early init");
         self.create_dir("/dev", Some(0o0755)).await?;
         self.create_dir("/proc", None).await?;
-        self.create_dir("/sys", None).await?;
+        self.create_dir("/sys", Some(0o0555)).await?;
         self.create_dir("/root", Some(0o0700)).await?;
         self.create_dir("/tmp", None).await?;
-        self.mount_kernel_fs("devtmpfs", "/dev", "mode=0755")
+        self.create_dir("/run", Some(0o0755)).await?;
+        self.mount_kernel_fs("devtmpfs", "/dev", "mode=0755", None)
+            .await?;
+        self.mount_kernel_fs("proc", "/proc", "", None).await?;
+        self.mount_kernel_fs("sysfs", "/sys", "", None).await?;
+        fs::symlink("/proc/self/fd", "/dev/fd").await?;
+        fs::symlink("/proc/self/fd/0", "/dev/stdin").await?;
+        fs::symlink("/proc/self/fd/1", "/dev/stdout").await?;
+        fs::symlink("/proc/self/fd/2", "/dev/stderr").await?;
+        self.mount_kernel_fs("cgroup2", "/sys/fs/cgroup", "", Some(MountFlags::RELATIME))
             .await?;
-        self.mount_kernel_fs("proc", "/proc", "").await?;
-        self.mount_kernel_fs("sysfs", "/sys", "").await?;
-        symlink("/proc/self/fd", "/dev/fd")?;
         Ok(())
     }
 
@@ -137,16 +164,19 @@ impl GuestInit {
         Ok(())
     }
 
-    async fn mount_kernel_fs(&mut self, fstype: &str, path: &str, data: &str) -> Result<()> {
-        let metadata = fs::metadata(path).await?;
-        if metadata.st_dev() == fs::metadata("/").await?.st_dev() {
-            trace!("mounting kernel fs {} to {}", fstype, path);
-            Mount::builder()
-                .fstype(FilesystemType::Manual(fstype))
-                .flags(MountFlags::NOEXEC | MountFlags::NOSUID)
-                .data(data)
-                .mount(fstype, path)?;
-        }
+    async fn mount_kernel_fs(
+        &mut self,
+        fstype: &str,
+        path: &str,
+        data: &str,
+        flags: Option<MountFlags>,
+    ) -> Result<()> {
+        trace!("mounting kernel fs {} to {}", fstype, path);
+        Mount::builder()
+            .fstype(FilesystemType::Manual(fstype))
+            .flags(MountFlags::NOEXEC | MountFlags::NOSUID | flags.unwrap_or(MountFlags::empty()))
+            .data(data)
+            .mount(fstype, path)?;
         Ok(())
     }
 
@@ -158,24 +188,41 @@ impl GuestInit {
         Ok(())
     }
 
-    async fn mount_squashfs_images(&mut self) -> Result<()> {
-        trace!("mounting squashfs images");
-        let image_mount_path = Path::new(IMAGE_MOUNT_PATH);
+    async fn mount_config_image(&mut self) -> Result<()> {
+        trace!("mounting config image");
         let config_mount_path = Path::new(CONFIG_MOUNT_PATH);
-        self.mount_squashfs(Path::new(IMAGE_BLOCK_DEVICE_PATH), image_mount_path)
-            .await?;
-        self.mount_squashfs(Path::new(CONFIG_BLOCK_DEVICE_PATH), config_mount_path)
+        self.mount_image(
+            Path::new(CONFIG_BLOCK_DEVICE_PATH),
+            config_mount_path,
+            LaunchPackedFormat::Squashfs,
+        )
+        .await?;
+        Ok(())
+    }
+
+    async fn mount_root_image(&mut self, format: LaunchPackedFormat) -> Result<()> {
+        trace!("mounting root image");
+        let image_mount_path = Path::new(IMAGE_MOUNT_PATH);
+        self.mount_image(Path::new(IMAGE_BLOCK_DEVICE_PATH), image_mount_path, format)
             .await?;
         Ok(())
     }
 
-    async fn mount_squashfs(&mut self, from: &Path, to: &Path) -> Result<()> {
-        trace!("mounting squashfs image {:?} to {:?}", from, to);
+    async fn mount_image(
+        &mut self,
+        from: &Path,
+        to: &Path,
+        format: LaunchPackedFormat,
+    ) -> Result<()> {
+        trace!("mounting {:?} image {:?} to {:?}", format, from, to);
         if !to.is_dir() {
             fs::create_dir(to).await?;
         }
         Mount::builder()
-            .fstype(FilesystemType::Manual("squashfs"))
+            .fstype(FilesystemType::Manual(match format {
+                LaunchPackedFormat::Squashfs => "squashfs",
+                LaunchPackedFormat::Erofs => "erofs",
+            }))
             .flags(MountFlags::RDONLY)
             .mount(from, to)?;
         Ok(())
@@ -249,40 +296,6 @@ impl GuestInit {
         Ok(serde_json::from_str(&content)?)
     }
 
-    async fn nuke_initrd(&mut self) -> Result<()> {
-        trace!("nuking initrd");
-        let initrd_dev = fs::metadata("/").await?.st_dev();
-        for item in WalkDir::new("/")
-            .same_file_system(true)
-            .follow_links(false)
-            .contents_first(true)
-        {
-            if item.is_err() {
-                continue;
-            }
-
-            let item = item?;
-            let metadata = match item.metadata() {
-                Ok(value) => value,
-                Err(_) => continue,
-            };
-
-            if metadata.st_dev() != initrd_dev {
-                continue;
-            }
-
-            if metadata.is_symlink() || metadata.is_file() {
-                let _ = fs::remove_file(item.path()).await;
-                trace!("deleting file {:?}", item.path());
-            } else if metadata.is_dir() {
-                let _ = fs::remove_dir(item.path()).await;
-                trace!("deleting directory {:?}", item.path());
-            }
-        }
-        trace!("nuked initrd");
-        Ok(())
-    }
-
     async fn bind_new_root(&mut self) -> Result<()> {
         self.mount_move_subtree(Path::new(SYS_PATH), Path::new(NEW_ROOT_SYS_PATH))
             .await?;
@@ -302,7 +315,7 @@ impl GuestInit {
         Ok(())
     }
 
-    async fn network_setup(&mut self, network: &LaunchNetwork) -> Result<()> {
+    async fn network_setup(&mut self, cfg: &LaunchInfo, network: &LaunchNetwork) -> Result<()> {
         trace!("setting up network for link");
 
         let etc = PathBuf::from_str("/etc")?;
@@ -310,14 +323,33 @@ impl GuestInit {
             fs::create_dir(etc).await?;
         }
         let resolv = PathBuf::from_str("/etc/resolv.conf")?;
-        let mut lines = vec!["# krata resolver configuration".to_string()];
-        for nameserver in &network.resolver.nameservers {
-            lines.push(format!("nameserver {}", nameserver));
+
+        {
+            let mut lines = vec!["# krata resolver configuration".to_string()];
+            for nameserver in &network.resolver.nameservers {
+                lines.push(format!("nameserver {}", nameserver));
+            }
+
+            let mut conf = lines.join("\n");
+            conf.push('\n');
+            fs::write(resolv, conf).await?;
+        }
+
+        let hosts = PathBuf::from_str("/etc/hosts")?;
+        if let Some(ref hostname) = cfg.hostname {
+            let mut lines = if hosts.exists() {
+                fs::read_to_string(&hosts)
+                    .await?
+                    .lines()
+                    .map(|x| x.to_string())
+                    .collect::<Vec<_>>()
+            } else {
+                vec!["127.0.0.1 localhost".to_string()]
+            };
+            lines.push(format!("127.0.1.1 {}", hostname));
+            fs::write(&hosts, lines.join("\n") + "\n").await?;
         }
 
-        let mut conf = lines.join("\n");
-        conf.push('\n');
-        fs::write(resolv, conf).await?;
         self.network_configure_ethtool(network).await?;
         self.network_configure_link(network).await?;
         Ok(())
@@ -327,6 +359,14 @@ impl GuestInit {
         let (connection, handle, _) = rtnetlink::new_connection()?;
         tokio::spawn(connection);
 
+        let mut links = handle.link().get().match_name("lo".to_string()).execute();
+        let Some(link) = links.try_next().await? else {
+            warn!("unable to find link named lo");
+            return Ok(());
+        };
+
+        handle.link().set(link.header.index).up().execute().await?;
+
         let ipv4_network: IpNetwork = network.ipv4.address.parse()?;
         let ipv4_gateway: Ipv4Addr = network.ipv4.gateway.parse()?;
         let ipv6_network: IpNetwork = network.ipv6.address.parse()?;
@@ -399,14 +439,19 @@ impl GuestInit {
         Ok(())
     }
 
-    async fn run(&mut self, config: &Config, launch: &LaunchInfo, idm: IdmClient) -> Result<()> {
+    async fn run(
+        &mut self,
+        config: &Config,
+        launch: &LaunchInfo,
+        idm: IdmInternalClient,
+    ) -> Result<()> {
         let mut cmd = match config.cmd() {
             None => vec![],
             Some(value) => value.clone(),
         };
 
         if launch.run.is_some() {
-            cmd = launch.run.as_ref().unwrap().clone();
+            cmd.clone_from(launch.run.as_ref().unwrap());
         }
 
         if let Some(entrypoint) = config.entrypoint() {
@@ -427,9 +472,14 @@ impl GuestInit {
         }
         env.extend(launch.env.clone());
         env.insert("KRATA_CONTAINER".to_string(), "1".to_string());
-        env.insert("TERM".to_string(), "vt100".to_string());
 
-        let path = GuestInit::resolve_executable(&env, path.into())?;
+        // If we were not provided a terminal definition in our launch manifest, we
+        // default to xterm as most terminal emulators support the xterm control codes.
+        if !env.contains_key("TERM") {
+            env.insert("TERM".to_string(), "xterm".to_string());
+        }
+
+        let path = resolve_executable(&env, path.into())?;
         let Some(file_name) = path.file_name() else {
             return Err(anyhow!("cannot get file name of command path"));
         };
@@ -454,10 +504,21 @@ impl GuestInit {
             working_dir = "/".to_string();
         }
 
-        self.fork_and_exec(idm, working_dir, path, cmd, env).await?;
+        let cgroup = self.init_cgroup().await?;
+        self.fork_and_exec(idm, cgroup, working_dir, path, cmd, env)
+            .await?;
         Ok(())
     }
 
+    async fn init_cgroup(&self) -> Result<Cgroup> {
+        trace!("initializing cgroup");
+        let hierarchy = cgroups_rs::hierarchies::auto();
+        let cgroup = Cgroup::new(hierarchy, "krata-guest-task")?;
+        cgroup.set_cgroup_type("threaded")?;
+        trace!("initialized cgroup");
+        Ok(cgroup)
+    }
+
     fn strings_as_cstrings(values: Vec<String>) -> Result<Vec<CString>> {
         let mut results: Vec<CString> = vec![];
         for value in values {
@@ -476,27 +537,6 @@ impl GuestInit {
         map
     }
 
-    fn resolve_executable(env: &HashMap<String, String>, path: PathBuf) -> Result<PathBuf> {
-        if path.is_absolute() {
-            return Ok(path);
-        }
-
-        if path.is_file() {
-            return Ok(path.absolutize()?.to_path_buf());
-        }
-
-        if let Some(path_var) = env.get("PATH") {
-            for item in path_var.split(':') {
-                let mut exe_path: PathBuf = item.into();
-                exe_path.push(&path);
-                if exe_path.is_file() {
-                    return Ok(exe_path);
-                }
-            }
-        }
-        Ok(path)
-    }
-
     fn env_list(env: HashMap<String, String>) -> Vec<String> {
         env.iter()
             .map(|(key, value)| format!("{}={}", key, value))
@@ -505,20 +545,22 @@ impl GuestInit {
 
     async fn fork_and_exec(
         &mut self,
-        idm: IdmClient,
+        idm: IdmInternalClient,
+        cgroup: Cgroup,
         working_dir: String,
         path: CString,
         cmd: Vec<CString>,
         env: Vec<CString>,
     ) -> Result<()> {
         match unsafe { fork()? } {
-            ForkResult::Parent { child } => self.background(idm, child).await,
-            ForkResult::Child => self.foreground(working_dir, path, cmd, env).await,
+            ForkResult::Parent { child } => self.background(idm, cgroup, child).await,
+            ForkResult::Child => self.foreground(cgroup, working_dir, path, cmd, env).await,
         }
     }
 
     async fn foreground(
         &mut self,
+        cgroup: Cgroup,
         working_dir: String,
         path: CString,
         cmd: Vec<CString>,
@@ -526,6 +568,7 @@ impl GuestInit {
     ) -> Result<()> {
         GuestInit::set_controlling_terminal()?;
         std::env::set_current_dir(working_dir)?;
+        cgroup.add_task(CgroupPid::from(std::process::id() as u64))?;
         execve(&path, &cmd, &env)?;
         Ok(())
     }
@@ -538,9 +581,35 @@ impl GuestInit {
         Ok(())
     }
 
-    async fn background(&mut self, idm: IdmClient, executed: Pid) -> Result<()> {
-        let mut background = GuestBackground::new(idm, executed).await?;
+    async fn background(
+        &mut self,
+        idm: IdmInternalClient,
+        cgroup: Cgroup,
+        executed: Pid,
+    ) -> Result<()> {
+        let mut background = GuestBackground::new(idm, cgroup, executed).await?;
         background.run().await?;
         Ok(())
     }
 }
+
+pub fn resolve_executable(env: &HashMap<String, String>, path: PathBuf) -> Result<PathBuf> {
+    if path.is_absolute() {
+        return Ok(path);
+    }
+
+    if path.is_file() {
+        return Ok(path.absolutize()?.to_path_buf());
+    }
+
+    if let Some(path_var) = env.get("PATH") {
+        for item in path_var.split(':') {
+            let mut exe_path: PathBuf = item.into();
+            exe_path.push(&path);
+            if exe_path.is_file() {
+                return Ok(exe_path);
+            }
+        }
+    }
+    Ok(path)
+}

crates/guest/src/lib.rs

@@ -6,7 +6,9 @@ use xenstore::{XsdClient, XsdInterface};
 
 pub mod background;
 pub mod childwait;
+pub mod exec;
 pub mod init;
+pub mod metrics;
 
 pub async fn death(code: c_int) -> Result<()> {
     let store = XsdClient::open().await?;

crates/guest/src/metrics.rs

@@ -0,0 +1,118 @@
+use std::{ops::Add, path::Path};
+
+use anyhow::Result;
+use krata::idm::internal::{MetricFormat, MetricNode};
+use sysinfo::Process;
+
+pub struct MetricsCollector {}
+
+impl MetricsCollector {
+    pub fn new() -> Result<Self> {
+        Ok(MetricsCollector {})
+    }
+
+    pub fn collect(&self) -> Result<MetricNode> {
+        let mut sysinfo = sysinfo::System::new();
+        Ok(MetricNode::structural(
+            "guest",
+            vec![
+                self.collect_system(&mut sysinfo)?,
+                self.collect_processes(&mut sysinfo)?,
+            ],
+        ))
+    }
+
+    fn collect_system(&self, sysinfo: &mut sysinfo::System) -> Result<MetricNode> {
+        sysinfo.refresh_memory();
+        Ok(MetricNode::structural(
+            "system",
+            vec![MetricNode::structural(
+                "memory",
+                vec![
+                    MetricNode::value("total", sysinfo.total_memory(), MetricFormat::Bytes),
+                    MetricNode::value("used", sysinfo.used_memory(), MetricFormat::Bytes),
+                    MetricNode::value("free", sysinfo.free_memory(), MetricFormat::Bytes),
+                ],
+            )],
+        ))
+    }
+
+    fn collect_processes(&self, sysinfo: &mut sysinfo::System) -> Result<MetricNode> {
+        sysinfo.refresh_processes();
+        let mut processes = Vec::new();
+        let mut sysinfo_processes = sysinfo.processes().values().collect::<Vec<_>>();
+        sysinfo_processes.sort_by_key(|x| x.pid());
+        for process in sysinfo_processes {
+            if process.thread_kind().is_some() {
+                continue;
+            }
+            processes.push(MetricsCollector::process_node(process)?);
+        }
+        Ok(MetricNode::structural("process", processes))
+    }
+
+    fn process_node(process: &Process) -> Result<MetricNode> {
+        let mut metrics = vec![];
+
+        if let Some(parent) = process.parent() {
+            metrics.push(MetricNode::value(
+                "parent",
+                parent.as_u32() as u64,
+                MetricFormat::Integer,
+            ));
+        }
+
+        if let Some(exe) = process.exe().and_then(path_as_str) {
+            metrics.push(MetricNode::raw_value("executable", exe));
+        }
+
+        if let Some(working_directory) = process.cwd().and_then(path_as_str) {
+            metrics.push(MetricNode::raw_value("cwd", working_directory));
+        }
+
+        let cmdline = process.cmd().to_vec();
+        metrics.push(MetricNode::raw_value("cmdline", cmdline));
+        metrics.push(MetricNode::structural(
+            "memory",
+            vec![
+                MetricNode::value("resident", process.memory(), MetricFormat::Bytes),
+                MetricNode::value("virtual", process.virtual_memory(), MetricFormat::Bytes),
+            ],
+        ));
+
+        metrics.push(MetricNode::value(
+            "lifetime",
+            process.run_time(),
+            MetricFormat::DurationSeconds,
+        ));
+        metrics.push(MetricNode::value(
+            "uid",
+            process.user_id().map(|x| (*x).add(0)).unwrap_or(0) as f64,
+            MetricFormat::Integer,
+        ));
+        metrics.push(MetricNode::value(
+            "gid",
+            process.group_id().map(|x| (*x).add(0)).unwrap_or(0) as f64,
+            MetricFormat::Integer,
+        ));
+        metrics.push(MetricNode::value(
+            "euid",
+            process
+                .effective_user_id()
+                .map(|x| (*x).add(0))
+                .unwrap_or(0) as f64,
+            MetricFormat::Integer,
+        ));
+        metrics.push(MetricNode::value(
+            "egid",
+            process.effective_group_id().map(|x| x.add(0)).unwrap_or(0) as f64,
+            MetricFormat::Integer,
+        ));
+
+        Ok(MetricNode::structural(process.pid().to_string(), metrics))
+    }
+}
+
+fn path_as_str(path: &Path) -> Option<String> {
+    String::from_utf8(path.as_os_str().as_encoded_bytes().to_vec()).ok()
+}

crates/krata/Cargo.toml

@@ -2,7 +2,7 @@
 name = "krata"
 description = "Client library and common services for the krata hypervisor."
 license.workspace = true
-version= "0.0.3"
+version.workspace = true
 homepage.workspace = true
 repository.workspace = true
 edition = "2021"
@@ -10,12 +10,15 @@ resolver = "2"
 
 [dependencies]
 anyhow = { workspace = true }
+async-trait = { workspace = true }
 bytes = { workspace = true }
 libc = { workspace = true }
 log = { workspace = true }
 once_cell = { workspace = true }
 prost = { workspace = true }
 prost-reflect = { workspace = true }
+prost-types = { workspace = true }
+scopeguard = { workspace = true }
 serde = { workspace = true }
 tonic = { workspace = true }
 tokio = { workspace = true }

crates/krata/build.rs

@@ -8,7 +8,8 @@ fn main() -> Result<()> {
             &mut config,
             &[
                 "proto/krata/v1/control.proto",
-                "proto/krata/internal/idm.proto",
+                "proto/krata/idm/transport.proto",
+                "proto/krata/idm/internal.proto",
             ],
             &["proto/"],
         )?;
@@ -16,7 +17,8 @@ fn main() -> Result<()> {
         config,
         &[
             "proto/krata/v1/control.proto",
-            "proto/krata/internal/idm.proto",
+            "proto/krata/idm/transport.proto",
+            "proto/krata/idm/internal.proto",
         ],
         &["proto/"],
     )?;

crates/krata/proto/krata/idm/internal.proto

@@ -0,0 +1,89 @@
+syntax = "proto3";
+
+package krata.idm.internal;
+
+option java_multiple_files = true;
+option java_package = "dev.krata.proto.idm.internal";
+option java_outer_classname = "IdmInternalProto";
+
+import "google/protobuf/struct.proto";
+
+message ExitEvent {
+    int32 code = 1;
+}
+
+message PingRequest {}
+
+message PingResponse {}
+
+message MetricsRequest {}
+
+message MetricsResponse {
+    MetricNode root = 1;
+}
+
+message MetricNode {
+    string name = 1;
+    google.protobuf.Value value = 2;
+    MetricFormat format = 3;
+    repeated MetricNode children = 4;
+}
+
+enum MetricFormat {
+    METRIC_FORMAT_UNKNOWN = 0;
+    METRIC_FORMAT_BYTES = 1;
+    METRIC_FORMAT_INTEGER = 2;
+    METRIC_FORMAT_DURATION_SECONDS = 3;
+}
+
+message ExecEnvVar {
+    string key = 1;
+    string value = 2;
+}
+
+message ExecStreamRequestStart {
+    repeated ExecEnvVar environment = 1;
+    repeated string command = 2;
+    string working_directory = 3;
+}
+
+message ExecStreamRequestStdin {
+    bytes data = 1;
+}
+
+message ExecStreamRequestUpdate {
+    oneof update {
+        ExecStreamRequestStart start = 1;
+        ExecStreamRequestStdin stdin = 2;
+    }
+}
+
+message ExecStreamResponseUpdate {
+    bool exited = 1;
+    string error = 2;
+    int32 exit_code = 3;
+    bytes stdout = 4;
+    bytes stderr = 5;
+}
+
+message Event {
+    oneof event {
+        ExitEvent exit = 1;
+    }
+}
+
+message Request {
+    oneof request {
+        PingRequest ping = 1;
+        MetricsRequest metrics = 2;
+        ExecStreamRequestUpdate exec_stream = 3;
+    }
+}
+
+message Response {
+    oneof response {
+        PingResponse ping = 1;
+        MetricsResponse metrics = 2;
+        ExecStreamResponseUpdate exec_stream = 3;
+    }
+}

crates/krata/proto/krata/idm/transport.proto

@@ -0,0 +1,27 @@
+syntax = "proto3";
+
+package krata.idm.transport;
+
+option java_multiple_files = true;
+option java_package = "dev.krata.proto.idm.transport";
+option java_outer_classname = "IdmTransportProto";
+
+message IdmTransportPacket {
+    uint64 id = 1;
+    uint64 channel = 2;
+    IdmTransportPacketForm form = 3;
+    bytes data = 4;
+}
+
+enum IdmTransportPacketForm {
+    IDM_TRANSPORT_PACKET_FORM_UNKNOWN = 0;
+    IDM_TRANSPORT_PACKET_FORM_RAW = 1;
+    IDM_TRANSPORT_PACKET_FORM_EVENT = 2;
+    IDM_TRANSPORT_PACKET_FORM_REQUEST = 3;
+    IDM_TRANSPORT_PACKET_FORM_RESPONSE = 4;
+    IDM_TRANSPORT_PACKET_FORM_STREAM_REQUEST = 5;
+    IDM_TRANSPORT_PACKET_FORM_STREAM_REQUEST_UPDATE = 6;
+    IDM_TRANSPORT_PACKET_FORM_STREAM_RESPONSE_UPDATE = 7;
+    IDM_TRANSPORT_PACKET_FORM_STREAM_REQUEST_CLOSED = 8;
+    IDM_TRANSPORT_PACKET_FORM_STREAM_RESPONSE_CLOSED = 9;
+}

crates/krata/proto/krata/internal/idm.proto

@@ -1,21 +0,0 @@
-syntax = "proto3";
-
-package krata.internal.idm;
-
-option java_multiple_files = true;
-option java_package = "dev.krata.proto.internal.idm";
-option java_outer_classname = "IdmProto";
-
-message IdmExitEvent {
-    int32 code = 1;
-}
-
-message IdmEvent {
-    oneof event {
-        IdmExitEvent exit = 1;
-    }
-}
-
-message IdmPacket {
-    IdmEvent event = 1;
-}

crates/krata/proto/krata/v1/common.proto

@@ -6,6 +6,8 @@ option java_multiple_files = true;
 option java_package = "dev.krata.proto.v1.common";
 option java_outer_classname = "CommonProto";
 
+import "google/protobuf/struct.proto";
+
 message Guest {
     string id = 1;
     GuestSpec spec = 2;
@@ -15,10 +17,14 @@ message Guest {
 message GuestSpec {
     string name = 1;
     GuestImageSpec image = 2;
-    uint32 vcpus = 3;
-    uint64 mem = 4;
-    GuestTaskSpec task = 5;
-    repeated GuestSpecAnnotation annotations = 6;
+    // If not specified, defaults to the daemon default kernel.
+    GuestImageSpec kernel = 3;
+    // If not specified, defaults to the daemon default initrd.
+    GuestImageSpec initrd = 4;
+    uint32 vcpus = 5;
+    uint64 mem = 6;
+    GuestTaskSpec task = 7;
+    repeated GuestSpecAnnotation annotations = 8;
 }
 
 message GuestImageSpec {
@@ -27,13 +33,23 @@ message GuestImageSpec {
     }
 }
 
+enum OciImageFormat {
+    OCI_IMAGE_FORMAT_UNKNOWN = 0;
+    OCI_IMAGE_FORMAT_SQUASHFS = 1;
+    OCI_IMAGE_FORMAT_EROFS = 2;
+    // Tar format is not launchable, and is intended for kernel images.
+    OCI_IMAGE_FORMAT_TAR = 3;
+}
+
 message GuestOciImageSpec {
-    string image = 1;
+    string digest = 1;
+    OciImageFormat format = 2;
 }
 
 message GuestTaskSpec {
     repeated GuestTaskSpecEnvVar environment = 1;
     repeated string command = 2;
+    string working_directory = 3;
 }
 
 message GuestTaskSpecEnvVar {
@@ -51,7 +67,8 @@ message GuestState {
     GuestNetworkState network = 2;
     GuestExitInfo exit_info = 3;
     GuestErrorInfo error_info = 4;
-    uint32 domid = 5;
+    string host = 5;
+    uint32 domid = 6;
 }
 
 enum GuestStatus {
@@ -80,3 +97,17 @@ message GuestExitInfo {
 message GuestErrorInfo {
     string message = 1;
 }
+
+message GuestMetricNode {
+    string name = 1;
+    google.protobuf.Value value = 2;
+    GuestMetricFormat format = 3;
+    repeated GuestMetricNode children = 4;
+}
+
+enum GuestMetricFormat {
+    GUEST_METRIC_FORMAT_UNKNOWN = 0;
+    GUEST_METRIC_FORMAT_BYTES = 1;
+    GUEST_METRIC_FORMAT_INTEGER = 2;
+    GUEST_METRIC_FORMAT_DURATION_SECONDS = 3;
+}

crates/krata/proto/krata/v1/control.proto

@@ -6,15 +6,34 @@ option java_multiple_files = true;
 option java_package = "dev.krata.proto.v1.control";
 option java_outer_classname = "ControlProto";
 
+import "krata/idm/transport.proto";
 import "krata/v1/common.proto";
 
 service ControlService {
+    rpc IdentifyHost(IdentifyHostRequest) returns (IdentifyHostReply);
+
     rpc CreateGuest(CreateGuestRequest) returns (CreateGuestReply);
     rpc DestroyGuest(DestroyGuestRequest) returns (DestroyGuestReply);
     rpc ResolveGuest(ResolveGuestRequest) returns (ResolveGuestReply);
     rpc ListGuests(ListGuestsRequest) returns (ListGuestsReply);
+
+    rpc ExecGuest(stream ExecGuestRequest) returns (stream ExecGuestReply);
+
     rpc ConsoleData(stream ConsoleDataRequest) returns (stream ConsoleDataReply);
+    rpc ReadGuestMetrics(ReadGuestMetricsRequest) returns (ReadGuestMetricsReply);
+    
+    rpc SnoopIdm(SnoopIdmRequest) returns (stream SnoopIdmReply);
     rpc WatchEvents(WatchEventsRequest) returns (stream WatchEventsReply);
+
+    rpc PullImage(PullImageRequest) returns (stream PullImageReply);
+}
+
+message IdentifyHostRequest {}
+
+message IdentifyHostReply {
+    string host_uuid = 1;
+    uint32 host_domid = 2;
+    string krata_version = 3;
 }
 
 message CreateGuestRequest {
@@ -45,6 +64,20 @@ message ListGuestsReply {
     repeated krata.v1.common.Guest guests = 1;
 }
 
+message ExecGuestRequest {
+    string guest_id = 1;
+    krata.v1.common.GuestTaskSpec task = 2;
+    bytes data = 3;
+}
+
+message ExecGuestReply {
+    bool exited = 1;
+    string error = 2;
+    int32 exit_code = 3;
+    bytes stdout = 4;
+    bytes stderr = 5;
+}
+
 message ConsoleDataRequest {
     string guest_id = 1;
     bytes data = 2;
@@ -65,3 +98,92 @@ message WatchEventsReply {
 message GuestChangedEvent {
     krata.v1.common.Guest guest = 1;
 }
+
+message ReadGuestMetricsRequest {
+    string guest_id = 1;
+}
+
+message ReadGuestMetricsReply {
+    krata.v1.common.GuestMetricNode root = 1;
+}
+
+message SnoopIdmRequest {}
+
+message SnoopIdmReply {
+    string from = 1;
+    string to = 2;
+    krata.idm.transport.IdmTransportPacket packet = 3;
+}
+
+message ImageProgress {
+    ImageProgressPhase phase = 1;
+    repeated ImageProgressLayer layers = 2;
+    ImageProgressIndication indication = 3;
+}
+
+enum ImageProgressPhase {
+    IMAGE_PROGRESS_PHASE_UNKNOWN = 0;
+    IMAGE_PROGRESS_PHASE_STARTED = 1;
+    IMAGE_PROGRESS_PHASE_RESOLVING = 2;
+    IMAGE_PROGRESS_PHASE_RESOLVED = 3;
+    IMAGE_PROGRESS_PHASE_CONFIG_DOWNLOAD = 4;
+    IMAGE_PROGRESS_PHASE_LAYER_DOWNLOAD = 5;
+    IMAGE_PROGRESS_PHASE_ASSEMBLE = 6;
+    IMAGE_PROGRESS_PHASE_PACK = 7;
+    IMAGE_PROGRESS_PHASE_COMPLETE = 8;
+}
+
+message ImageProgressLayer {
+    string id = 1;
+    ImageProgressLayerPhase phase = 2;
+    ImageProgressIndication indication = 3;
+}
+
+enum ImageProgressLayerPhase {
+    IMAGE_PROGRESS_LAYER_PHASE_UNKNOWN = 0;
+    IMAGE_PROGRESS_LAYER_PHASE_WAITING = 1;
+    IMAGE_PROGRESS_LAYER_PHASE_DOWNLOADING = 2;
+    IMAGE_PROGRESS_LAYER_PHASE_DOWNLOADED = 3;
+    IMAGE_PROGRESS_LAYER_PHASE_EXTRACTING = 4;
+    IMAGE_PROGRESS_LAYER_PHASE_EXTRACTED = 5;
+}
+
+message ImageProgressIndication {
+    oneof indication {
+        ImageProgressIndicationBar bar = 1;
+        ImageProgressIndicationSpinner spinner = 2;
+        ImageProgressIndicationHidden hidden = 3;
+        ImageProgressIndicationCompleted completed = 4;
+    }
+}
+
+message ImageProgressIndicationBar {
+    string message = 1;
+    uint64 current = 2;
+    uint64 total = 3;
+    bool is_bytes = 4;
+}
+
+message ImageProgressIndicationSpinner {
+    string message = 1;
+}
+
+message ImageProgressIndicationHidden {}
+
+message ImageProgressIndicationCompleted {
+    string message = 1;
+    uint64 total = 2;
+    bool is_bytes = 3;
+}
+
+message PullImageRequest {
+    string image = 1;
+    krata.v1.common.OciImageFormat format = 2;
+    bool overwrite_cache = 3;
+}
+
+message PullImageReply {
+    ImageProgress progress = 1;
+    string digest = 2;
+    krata.v1.common.OciImageFormat format = 3;
+}

crates/krata/src/events.rs

@@ -1,11 +1,14 @@
-use std::sync::Arc;
+use std::{sync::Arc, time::Duration};
 
-use crate::v1::control::{watch_events_reply::Event, WatchEventsReply};
+use crate::v1::control::{
+    control_service_client::ControlServiceClient, watch_events_reply::Event, WatchEventsReply,
+    WatchEventsRequest,
+};
 use anyhow::Result;
-use log::trace;
-use tokio::{sync::broadcast, task::JoinHandle};
+use log::{error, trace, warn};
+use tokio::{sync::broadcast, task::JoinHandle, time::sleep};
 use tokio_stream::StreamExt;
-use tonic::Streaming;
+use tonic::{transport::Channel, Streaming};
 
 #[derive(Clone)]
 pub struct EventStream {
@@ -14,27 +17,12 @@ pub struct EventStream {
 }
 
 impl EventStream {
-    pub async fn open(mut events: Streaming<WatchEventsReply>) -> Result<Self> {
+    pub async fn open(client: ControlServiceClient<Channel>) -> Result<Self> {
         let (sender, _) = broadcast::channel(1000);
         let emit = sender.clone();
         let task = tokio::task::spawn(async move {
-            loop {
-                let Some(result) = events.next().await else {
-                    break;
-                };
-
-                let reply = match result {
-                    Ok(reply) => reply,
-                    Err(error) => {
-                        trace!("event stream processing failed: {}", error);
-                        break;
-                    }
-                };
-
-                let Some(event) = reply.event else {
-                    continue;
-                };
-                let _ = emit.send(event);
+            if let Err(error) = EventStream::process(client, emit).await {
+                error!("failed to process event stream: {}", error);
             }
         });
         Ok(Self {
@@ -43,6 +31,48 @@ impl EventStream {
         })
     }
 
+    async fn process(
+        mut client: ControlServiceClient<Channel>,
+        emit: broadcast::Sender<Event>,
+    ) -> Result<()> {
+        let mut events: Option<Streaming<WatchEventsReply>> = None;
+        loop {
+            let mut stream = match events {
+                Some(stream) => stream,
+                None => {
+                    let result = client.watch_events(WatchEventsRequest {}).await;
+                    if let Err(error) = result {
+                        warn!("failed to watch events: {}", error);
+                        sleep(Duration::from_secs(1)).await;
+                        continue;
+                    }
+                    result.unwrap().into_inner()
+                }
+            };
+
+            let Some(result) = stream.next().await else {
+                events = None;
+                continue;
+            };
+
+            let reply = match result {
+                Ok(reply) => reply,
+                Err(error) => {
+                    trace!("event stream processing failed: {}", error);
+                    events = None;
+                    continue;
+                }
+            };
+
+            let Some(event) = reply.event else {
+                events = Some(stream);
+                continue;
+            };
+            let _ = emit.send(event);
+            events = Some(stream);
+        }
+    }
+
     pub fn subscribe(&self) -> broadcast::Receiver<Event> {
         self.sender.subscribe()
     }

crates/krata/src/idm/client.rs

@@ -1,53 +1,63 @@
-use std::path::Path;
+use std::{
+    collections::HashMap,
+    path::Path,
+    sync::{
+        atomic::{AtomicBool, Ordering},
+        Arc,
+    },
+    time::Duration,
+};
 
-use super::protocol::IdmPacket;
 use anyhow::{anyhow, Result};
-use bytes::BytesMut;
-use log::error;
+use log::{debug, error};
 use nix::sys::termios::{cfmakeraw, tcgetattr, tcsetattr, SetArg};
 use prost::Message;
 use tokio::{
     fs::File,
     io::{unix::AsyncFd, AsyncReadExt, AsyncWriteExt},
     select,
-    sync::mpsc::{channel, Receiver, Sender},
+    sync::{
+        broadcast,
+        mpsc::{self, Receiver, Sender},
+        oneshot, Mutex,
+    },
     task::JoinHandle,
+    time::timeout,
 };
 
+use super::{
+    internal,
+    serialize::{IdmRequest, IdmSerializable},
+    transport::{IdmTransportPacket, IdmTransportPacketForm},
+};
+
+type OneshotRequestMap<R> = Arc<Mutex<HashMap<u64, oneshot::Sender<<R as IdmRequest>::Response>>>>;
+type StreamRequestMap<R> = Arc<Mutex<HashMap<u64, Sender<<R as IdmRequest>::Response>>>>;
+type StreamRequestUpdateMap<R> = Arc<Mutex<HashMap<u64, mpsc::Sender<R>>>>;
+pub type IdmInternalClient = IdmClient<internal::Request, internal::Event>;
+
 const IDM_PACKET_QUEUE_LEN: usize = 100;
+const IDM_REQUEST_TIMEOUT_SECS: u64 = 10;
+const IDM_PACKET_MAX_SIZE: usize = 20 * 1024 * 1024;
 
-pub struct IdmClient {
-    pub receiver: Receiver<IdmPacket>,
-    pub sender: Sender<IdmPacket>,
-    task: JoinHandle<()>,
+#[async_trait::async_trait]
+pub trait IdmBackend: Send {
+    async fn recv(&mut self) -> Result<IdmTransportPacket>;
+    async fn send(&mut self, packet: IdmTransportPacket) -> Result<()>;
 }
 
-impl Drop for IdmClient {
-    fn drop(&mut self) {
-        self.task.abort();
-    }
+pub struct IdmFileBackend {
+    read_fd: Arc<Mutex<AsyncFd<File>>>,
+    write: Arc<Mutex<File>>,
 }
 
-impl IdmClient {
-    pub async fn open<P: AsRef<Path>>(path: P) -> Result<IdmClient> {
-        let file = File::options()
-            .read(true)
-            .write(true)
-            .create(false)
-            .open(path)
-            .await?;
-        IdmClient::set_raw_port(&file)?;
-        let (rx_sender, rx_receiver) = channel(IDM_PACKET_QUEUE_LEN);
-        let (tx_sender, tx_receiver) = channel(IDM_PACKET_QUEUE_LEN);
-        let task = tokio::task::spawn(async move {
-            if let Err(error) = IdmClient::process(file, rx_sender, tx_receiver).await {
-                error!("failed to handle idm client processing: {}", error);
-            }
-        });
-        Ok(IdmClient {
-            receiver: rx_receiver,
-            sender: tx_sender,
-            task,
+impl IdmFileBackend {
+    pub async fn new(read_file: File, write_file: File) -> Result<IdmFileBackend> {
+        IdmFileBackend::set_raw_port(&read_file)?;
+        IdmFileBackend::set_raw_port(&write_file)?;
+        Ok(IdmFileBackend {
+            read_fd: Arc::new(Mutex::new(AsyncFd::new(read_file)?)),
+            write: Arc::new(Mutex::new(write_file)),
         })
     }
 
@@ -57,31 +67,413 @@ impl IdmClient {
         tcsetattr(file, SetArg::TCSANOW, &termios)?;
         Ok(())
     }
+}
 
+#[async_trait::async_trait]
+impl IdmBackend for IdmFileBackend {
+    async fn recv(&mut self) -> Result<IdmTransportPacket> {
+        let mut fd = self.read_fd.lock().await;
+        let mut guard = fd.readable_mut().await?;
+        let b1 = guard.get_inner_mut().read_u8().await?;
+        if b1 != 0xff {
+            return Ok(IdmTransportPacket::default());
+        }
+        let b2 = guard.get_inner_mut().read_u8().await?;
+        if b2 != 0xff {
+            return Ok(IdmTransportPacket::default());
+        }
+        let size = guard.get_inner_mut().read_u32_le().await?;
+        if size == 0 {
+            return Ok(IdmTransportPacket::default());
+        }
+        let mut buffer = vec![0u8; size as usize];
+        guard.get_inner_mut().read_exact(&mut buffer).await?;
+        match IdmTransportPacket::decode(buffer.as_slice()) {
+            Ok(packet) => Ok(packet),
+            Err(error) => Err(anyhow!("received invalid idm packet: {}", error)),
+        }
+    }
+
+    async fn send(&mut self, packet: IdmTransportPacket) -> Result<()> {
+        let mut file = self.write.lock().await;
+        let data = packet.encode_to_vec();
+        file.write_all(&[0xff, 0xff]).await?;
+        file.write_u32_le(data.len() as u32).await?;
+        file.write_all(&data).await?;
+        Ok(())
+    }
+}
+
+#[derive(Clone)]
+pub struct IdmClient<R: IdmRequest, E: IdmSerializable> {
+    channel: u64,
+    request_backend_sender: broadcast::Sender<(u64, R)>,
+    request_stream_backend_sender: broadcast::Sender<IdmClientStreamResponseHandle<R>>,
+    next_request_id: Arc<Mutex<u64>>,
+    event_receiver_sender: broadcast::Sender<E>,
+    tx_sender: Sender<IdmTransportPacket>,
+    requests: OneshotRequestMap<R>,
+    request_streams: StreamRequestMap<R>,
+    task: Arc<JoinHandle<()>>,
+}
+
+impl<R: IdmRequest, E: IdmSerializable> Drop for IdmClient<R, E> {
+    fn drop(&mut self) {
+        if Arc::strong_count(&self.task) <= 1 {
+            self.task.abort();
+        }
+    }
+}
+
+pub struct IdmClientStreamRequestHandle<R: IdmRequest, E: IdmSerializable> {
+    pub id: u64,
+    pub receiver: Receiver<R::Response>,
+    pub client: IdmClient<R, E>,
+}
+
+impl<R: IdmRequest, E: IdmSerializable> IdmClientStreamRequestHandle<R, E> {
+    pub async fn update(&self, request: R) -> Result<()> {
+        self.client
+            .tx_sender
+            .send(IdmTransportPacket {
+                id: self.id,
+                channel: self.client.channel,
+                form: IdmTransportPacketForm::StreamRequestUpdate.into(),
+                data: request.encode()?,
+            })
+            .await?;
+        Ok(())
+    }
+}
+
+impl<R: IdmRequest, E: IdmSerializable> Drop for IdmClientStreamRequestHandle<R, E> {
+    fn drop(&mut self) {
+        let id = self.id;
+        let client = self.client.clone();
+        tokio::task::spawn(async move {
+            let _ = client
+                .tx_sender
+                .send(IdmTransportPacket {
+                    id,
+                    channel: client.channel,
+                    form: IdmTransportPacketForm::StreamRequestClosed.into(),
+                    data: vec![],
+                })
+                .await;
+        });
+    }
+}
+
+#[derive(Clone)]
+pub struct IdmClientStreamResponseHandle<R: IdmRequest> {
+    pub initial: R,
+    pub id: u64,
+    channel: u64,
+    tx_sender: Sender<IdmTransportPacket>,
+    receiver: Arc<Mutex<Option<Receiver<R>>>>,
+}
+
+impl<R: IdmRequest> IdmClientStreamResponseHandle<R> {
+    pub async fn respond(&self, response: R::Response) -> Result<()> {
+        self.tx_sender
+            .send(IdmTransportPacket {
+                id: self.id,
+                channel: self.channel,
+                form: IdmTransportPacketForm::StreamResponseUpdate.into(),
+                data: response.encode()?,
+            })
+            .await?;
+        Ok(())
+    }
+
+    pub async fn take(&self) -> Result<Receiver<R>> {
+        let mut guard = self.receiver.lock().await;
+        let Some(receiver) = (*guard).take() else {
+            return Err(anyhow!("request has already been claimed!"));
+        };
+        Ok(receiver)
+    }
+}
+
+impl<R: IdmRequest> Drop for IdmClientStreamResponseHandle<R> {
+    fn drop(&mut self) {
+        if Arc::strong_count(&self.receiver) <= 1 {
+            let id = self.id;
+            let channel = self.channel;
+            let tx_sender = self.tx_sender.clone();
+            tokio::task::spawn(async move {
+                let _ = tx_sender
+                    .send(IdmTransportPacket {
+                        id,
+                        channel,
+                        form: IdmTransportPacketForm::StreamResponseClosed.into(),
+                        data: vec![],
+                    })
+                    .await;
+            });
+        }
+    }
+}
+
+impl<R: IdmRequest, E: IdmSerializable> IdmClient<R, E> {
+    pub async fn new(channel: u64, backend: Box<dyn IdmBackend>) -> Result<Self> {
+        let requests = Arc::new(Mutex::new(HashMap::new()));
+        let request_streams = Arc::new(Mutex::new(HashMap::new()));
+        let request_update_streams = Arc::new(Mutex::new(HashMap::new()));
+        let (event_sender, event_receiver) = broadcast::channel(IDM_PACKET_QUEUE_LEN);
+        let (internal_request_backend_sender, _) = broadcast::channel(IDM_PACKET_QUEUE_LEN);
+        let (internal_request_stream_backend_sender, _) = broadcast::channel(IDM_PACKET_QUEUE_LEN);
+        let (tx_sender, tx_receiver) = mpsc::channel(IDM_PACKET_QUEUE_LEN);
+        let backend_event_sender = event_sender.clone();
+        let request_backend_sender = internal_request_backend_sender.clone();
+        let request_stream_backend_sender = internal_request_stream_backend_sender.clone();
+        let requests_for_client = requests.clone();
+        let request_streams_for_client = request_streams.clone();
+        let tx_sender_for_client = tx_sender.clone();
+        let task = tokio::task::spawn(async move {
+            if let Err(error) = IdmClient::process(
+                backend,
+                channel,
+                tx_sender,
+                backend_event_sender,
+                requests,
+                request_streams,
+                request_update_streams,
+                internal_request_backend_sender,
+                internal_request_stream_backend_sender,
+                event_receiver,
+                tx_receiver,
+            )
+            .await
+            {
+                debug!("failed to handle idm client processing: {}", error);
+            }
+        });
+        Ok(IdmClient {
+            channel,
+            next_request_id: Arc::new(Mutex::new(0)),
+            event_receiver_sender: event_sender.clone(),
+            request_backend_sender,
+            request_stream_backend_sender,
+            requests: requests_for_client,
+            request_streams: request_streams_for_client,
+            tx_sender: tx_sender_for_client,
+            task: Arc::new(task),
+        })
+    }
+
+    pub async fn open<P: AsRef<Path>>(channel: u64, path: P) -> Result<Self> {
+        let read_file = File::options()
+            .read(true)
+            .write(false)
+            .create(false)
+            .open(&path)
+            .await?;
+        let write_file = File::options()
+            .read(false)
+            .write(true)
+            .create(false)
+            .open(path)
+            .await?;
+        let backend = IdmFileBackend::new(read_file, write_file).await?;
+        IdmClient::new(channel, Box::new(backend) as Box<dyn IdmBackend>).await
+    }
+
+    pub async fn emit<T: IdmSerializable>(&self, event: T) -> Result<()> {
+        let id = {
+            let mut guard = self.next_request_id.lock().await;
+            let req = *guard;
+            *guard = req.wrapping_add(1);
+            req
+        };
+        self.tx_sender
+            .send(IdmTransportPacket {
+                id,
+                form: IdmTransportPacketForm::Event.into(),
+                channel: self.channel,
+                data: event.encode()?,
+            })
+            .await?;
+        Ok(())
+    }
+
+    pub async fn requests(&self) -> Result<broadcast::Receiver<(u64, R)>> {
+        Ok(self.request_backend_sender.subscribe())
+    }
+
+    pub async fn request_streams(
+        &self,
+    ) -> Result<broadcast::Receiver<IdmClientStreamResponseHandle<R>>> {
+        Ok(self.request_stream_backend_sender.subscribe())
+    }
+
+    pub async fn respond<T: IdmSerializable>(&self, id: u64, response: T) -> Result<()> {
+        let packet = IdmTransportPacket {
+            id,
+            form: IdmTransportPacketForm::Response.into(),
+            channel: self.channel,
+            data: response.encode()?,
+        };
+        self.tx_sender.send(packet).await?;
+        Ok(())
+    }
+
+    pub async fn subscribe(&self) -> Result<broadcast::Receiver<E>> {
+        Ok(self.event_receiver_sender.subscribe())
+    }
+
+    pub async fn send(&self, request: R) -> Result<R::Response> {
+        let (sender, receiver) = oneshot::channel::<R::Response>();
+        let req = {
+            let mut guard = self.next_request_id.lock().await;
+            let req = *guard;
+            *guard = req.wrapping_add(1);
+            req
+        };
+        let mut requests = self.requests.lock().await;
+        requests.insert(req, sender);
+        drop(requests);
+        let success = AtomicBool::new(false);
+        let _guard = scopeguard::guard(self.requests.clone(), |requests| {
+            if success.load(Ordering::Acquire) {
+                return;
+            }
+            tokio::task::spawn(async move {
+                let mut requests = requests.lock().await;
+                requests.remove(&req);
+            });
+        });
+        self.tx_sender
+            .send(IdmTransportPacket {
+                id: req,
+                channel: self.channel,
+                form: IdmTransportPacketForm::Request.into(),
+                data: request.encode()?,
+            })
+            .await?;
+
+        let response = timeout(Duration::from_secs(IDM_REQUEST_TIMEOUT_SECS), receiver).await??;
+        success.store(true, Ordering::Release);
+        Ok(response)
+    }
+
+    pub async fn send_stream(&self, request: R) -> Result<IdmClientStreamRequestHandle<R, E>> {
+        let (sender, receiver) = mpsc::channel::<R::Response>(100);
+        let req = {
+            let mut guard = self.next_request_id.lock().await;
+            let req = *guard;
+            *guard = req.wrapping_add(1);
+            req
+        };
+        let mut requests = self.request_streams.lock().await;
+        requests.insert(req, sender);
+        drop(requests);
+        self.tx_sender
+            .send(IdmTransportPacket {
+                id: req,
+                channel: self.channel,
+                form: IdmTransportPacketForm::StreamRequest.into(),
+                data: request.encode()?,
+            })
+            .await?;
+        Ok(IdmClientStreamRequestHandle {
+            id: req,
+            receiver,
+            client: self.clone(),
+        })
+    }
+
+    #[allow(clippy::too_many_arguments)]
     async fn process(
-        file: File,
-        sender: Sender<IdmPacket>,
-        mut receiver: Receiver<IdmPacket>,
+        mut backend: Box<dyn IdmBackend>,
+        channel: u64,
+        tx_sender: Sender<IdmTransportPacket>,
+        event_sender: broadcast::Sender<E>,
+        requests: OneshotRequestMap<R>,
+        request_streams: StreamRequestMap<R>,
+        request_update_streams: StreamRequestUpdateMap<R>,
+        request_backend_sender: broadcast::Sender<(u64, R)>,
+        request_stream_backend_sender: broadcast::Sender<IdmClientStreamResponseHandle<R>>,
+        _event_receiver: broadcast::Receiver<E>,
+        mut receiver: Receiver<IdmTransportPacket>,
     ) -> Result<()> {
-        let mut file = AsyncFd::new(file)?;
         loop {
             select! {
-                x = file.readable_mut() => match x {
-                    Ok(mut guard) => {
-                        let size = guard.get_inner_mut().read_u16_le().await?;
-                        if size == 0 {
+                x = backend.recv() => match x {
+                    Ok(packet) => {
+                        if packet.channel != channel {
                             continue;
                         }
-                        let mut buffer = BytesMut::with_capacity(size as usize);
-                        guard.get_inner_mut().read_exact(&mut buffer).await?;
-                        match IdmPacket::decode(buffer) {
-                            Ok(packet) => {
-                                sender.send(packet).await?;
+
+                        match packet.form() {
+                            IdmTransportPacketForm::Event => {
+                                if let Ok(event) = E::decode(&packet.data) {
+                                    let _ = event_sender.send(event);
+                                }
                             },
 
-                            Err(error) => {
-                                error!("received invalid idm packet: {}", error);
+                            IdmTransportPacketForm::Request => {
+                                if let Ok(request) = R::decode(&packet.data) {
+                                    let _ = request_backend_sender.send((packet.id, request));
+                                }
+                            },
+
+                            IdmTransportPacketForm::Response => {
+                                let mut requests = requests.lock().await;
+                                if let Some(sender) = requests.remove(&packet.id) {
+                                    drop(requests);
+
+                                    if let Ok(response) = R::Response::decode(&packet.data) {
+                                        let _ = sender.send(response);
+                                    }
+                                }
+                            },
+
+                            IdmTransportPacketForm::StreamRequest => {
+                                if let Ok(request) = R::decode(&packet.data) {
+                                    let mut update_streams = request_update_streams.lock().await;
+                                    let (sender, receiver) = mpsc::channel(100);
+                                    update_streams.insert(packet.id, sender.clone());
+                                    let handle = IdmClientStreamResponseHandle {
+                                        initial: request,
+                                        id: packet.id,
+                                        channel,
+                                        tx_sender: tx_sender.clone(),
+                                        receiver: Arc::new(Mutex::new(Some(receiver))),
+                                    };
+                                    let _ = request_stream_backend_sender.send(handle);
+                                }
                             }
+
+                            IdmTransportPacketForm::StreamRequestUpdate => {
+                                if let Ok(request) = R::decode(&packet.data) {
+                                    let mut update_streams = request_update_streams.lock().await;
+                                    if let Some(stream) = update_streams.get_mut(&packet.id) {
+                                        let _ = stream.try_send(request);
+                                    }
+                                }
+                            }
+
+                            IdmTransportPacketForm::StreamRequestClosed => {
+                                let mut update_streams = request_update_streams.lock().await;
+                                update_streams.remove(&packet.id);
+                            }
+
+                            IdmTransportPacketForm::StreamResponseUpdate => {
+                                let requests = request_streams.lock().await;
+                                if let Some(sender) = requests.get(&packet.id) {
+                                    if let Ok(response) = R::Response::decode(&packet.data) {
+                                        let _ = sender.try_send(response);
+                                    }
+                                }
+                            }
+
+                            IdmTransportPacketForm::StreamResponseClosed => {
+                                let mut requests = request_streams.lock().await;
+                                requests.remove(&packet.id);
+                            }
+
+                            _ => {},
                         }
                     },
 
@@ -91,13 +483,12 @@ impl IdmClient {
                 },
                 x = receiver.recv() => match x {
                     Some(packet) => {
-                        let data = packet.encode_to_vec();
-                        if data.len() > u16::MAX as usize {
-                            error!("unable to send idm packet, packet size exceeded (tried to send {} bytes)", data.len());
+                        let length = packet.encoded_len();
+                        if length > IDM_PACKET_MAX_SIZE {
+                            error!("unable to send idm packet, packet size exceeded (tried to send {} bytes)", length);
                             continue;
                         }
-                        file.get_mut().write_u16_le(data.len() as u16).await?;
-                        file.get_mut().write_all(&data).await?;
+                        backend.send(packet).await?;
                     },
 
                     None => {

crates/krata/src/idm/internal.rs

@@ -0,0 +1,129 @@
+use anyhow::Result;
+use prost::Message;
+use prost_types::{ListValue, Value};
+
+use super::serialize::{IdmRequest, IdmSerializable};
+
+include!(concat!(env!("OUT_DIR"), "/krata.idm.internal.rs"));
+
+pub const INTERNAL_IDM_CHANNEL: u64 = 0;
+
+impl IdmSerializable for Event {
+    fn encode(&self) -> Result<Vec<u8>> {
+        Ok(self.encode_to_vec())
+    }
+
+    fn decode(bytes: &[u8]) -> Result<Self> {
+        Ok(<Self as prost::Message>::decode(bytes)?)
+    }
+}
+
+impl IdmSerializable for Request {
+    fn encode(&self) -> Result<Vec<u8>> {
+        Ok(self.encode_to_vec())
+    }
+
+    fn decode(bytes: &[u8]) -> Result<Self> {
+        Ok(<Self as prost::Message>::decode(bytes)?)
+    }
+}
+
+impl IdmRequest for Request {
+    type Response = Response;
+}
+
+impl IdmSerializable for Response {
+    fn encode(&self) -> Result<Vec<u8>> {
+        Ok(self.encode_to_vec())
+    }
+
+    fn decode(bytes: &[u8]) -> Result<Self> {
+        Ok(<Self as prost::Message>::decode(bytes)?)
+    }
+}
+
+pub trait AsIdmMetricValue {
+    fn as_metric_value(&self) -> Value;
+}
+
+impl MetricNode {
+    pub fn structural<N: AsRef<str>>(name: N, children: Vec<MetricNode>) -> MetricNode {
+        MetricNode {
+            name: name.as_ref().to_string(),
+            value: None,
+            format: MetricFormat::Unknown.into(),
+            children,
+        }
+    }
+
+    pub fn raw_value<N: AsRef<str>, V: AsIdmMetricValue>(name: N, value: V) -> MetricNode {
+        MetricNode {
+            name: name.as_ref().to_string(),
+            value: Some(value.as_metric_value()),
+            format: MetricFormat::Unknown.into(),
+            children: vec![],
+        }
+    }
+
+    pub fn value<N: AsRef<str>, V: AsIdmMetricValue>(
+        name: N,
+        value: V,
+        format: MetricFormat,
+    ) -> MetricNode {
+        MetricNode {
+            name: name.as_ref().to_string(),
+            value: Some(value.as_metric_value()),
+            format: format.into(),
+            children: vec![],
+        }
+    }
+}
+
+impl AsIdmMetricValue for String {
+    fn as_metric_value(&self) -> Value {
+        Value {
+            kind: Some(prost_types::value::Kind::StringValue(self.to_string())),
+        }
+    }
+}
+
+impl AsIdmMetricValue for &str {
+    fn as_metric_value(&self) -> Value {
+        Value {
+            kind: Some(prost_types::value::Kind::StringValue(self.to_string())),
+        }
+    }
+}
+
+impl AsIdmMetricValue for u64 {
+    fn as_metric_value(&self) -> Value {
+        numeric(*self as f64)
+    }
+}
+
+impl AsIdmMetricValue for i64 {
+    fn as_metric_value(&self) -> Value {
+        numeric(*self as f64)
+    }
+}
+
+impl AsIdmMetricValue for f64 {
+    fn as_metric_value(&self) -> Value {
+        numeric(*self)
+    }
+}
+
+impl<T: AsIdmMetricValue> AsIdmMetricValue for Vec<T> {
+    fn as_metric_value(&self) -> Value {
+        let values = self.iter().map(|x| x.as_metric_value()).collect::<_>();
+        Value {
+            kind: Some(prost_types::value::Kind::ListValue(ListValue { values })),
+        }
+    }
+}
+
+fn numeric(value: f64) -> Value {
+    Value {
+        kind: Some(prost_types::value::Kind::NumberValue(value)),
+    }
+}

crates/krata/src/idm/mod.rs

@@ -1,3 +1,5 @@
 #[cfg(unix)]
 pub mod client;
-pub mod protocol;
+pub mod internal;
+pub mod serialize;
+pub mod transport;

crates/krata/src/idm/protocol.rs

@@ -1 +0,0 @@
-include!(concat!(env!("OUT_DIR"), "/krata.internal.idm.rs"));

crates/krata/src/idm/serialize.rs

@@ -0,0 +1,10 @@
+use anyhow::Result;
+
+pub trait IdmSerializable: Sized + Clone + Send + Sync + 'static {
+    fn decode(bytes: &[u8]) -> Result<Self>;
+    fn encode(&self) -> Result<Vec<u8>>;
+}
+
+pub trait IdmRequest: IdmSerializable {
+    type Response: IdmSerializable;
+}

crates/krata/src/idm/transport.rs

@@ -0,0 +1 @@
+include!(concat!(env!("OUT_DIR"), "/krata.idm.transport.rs"));

crates/krata/src/launchcfg.rs

@@ -2,24 +2,30 @@ use std::collections::HashMap;
 
 use serde::{Deserialize, Serialize};
 
-#[derive(Serialize, Deserialize, Debug)]
+#[derive(Serialize, Deserialize, Debug, Clone)]
+pub enum LaunchPackedFormat {
+    Squashfs,
+    Erofs,
+}
+
+#[derive(Serialize, Deserialize, Debug, Clone)]
 pub struct LaunchNetworkIpv4 {
     pub address: String,
     pub gateway: String,
 }
 
-#[derive(Serialize, Deserialize, Debug)]
+#[derive(Serialize, Deserialize, Debug, Clone)]
 pub struct LaunchNetworkIpv6 {
     pub address: String,
     pub gateway: String,
 }
 
-#[derive(Serialize, Deserialize, Debug)]
+#[derive(Serialize, Deserialize, Debug, Clone)]
 pub struct LaunchNetworkResolver {
     pub nameservers: Vec<String>,
 }
 
-#[derive(Serialize, Deserialize, Debug)]
+#[derive(Serialize, Deserialize, Debug, Clone)]
 pub struct LaunchNetwork {
     pub link: String,
     pub ipv4: LaunchNetworkIpv4,
@@ -27,8 +33,15 @@ pub struct LaunchNetwork {
     pub resolver: LaunchNetworkResolver,
 }
 
-#[derive(Serialize, Deserialize, Debug)]
+#[derive(Serialize, Deserialize, Debug, Clone)]
+pub struct LaunchRoot {
+    pub format: LaunchPackedFormat,
+}
+
+#[derive(Serialize, Deserialize, Debug, Clone)]
 pub struct LaunchInfo {
+    pub root: LaunchRoot,
+    pub hostname: Option<String>,
     pub network: Option<LaunchNetwork>,
     pub env: HashMap<String, String>,
     pub run: Option<Vec<String>>,

crates/krata/src/v1/common.rs

@@ -1 +1,2 @@
+#![allow(clippy::all)]
 tonic::include_proto!("krata.v1.common");

crates/krata/src/v1/control.rs

@@ -1 +1,2 @@
+#![allow(clippy::all)]
 tonic::include_proto!("krata.v1.control");

crates/network/Cargo.toml

@@ -2,7 +2,7 @@
 name = "krata-network"
 description = "Networking services for the krata hypervisor."
 license.workspace = true
-version= "0.0.3"
+version.workspace = true
 homepage.workspace = true
 repository.workspace = true
 edition = "2021"
@@ -16,7 +16,7 @@ clap = { workspace = true }
 env_logger = { workspace = true }
 etherparse = { workspace = true }
 futures = { workspace = true }
-krata = { path = "../krata", version = "^0.0.3" }
+krata = { path = "../krata", version = "^0.0.10" }
 krata-advmac = { workspace = true }
 libc = { workspace = true }
 log = { workspace = true }

crates/network/src/autonet.rs

@@ -5,7 +5,7 @@ use krata::{
         common::Guest,
         control::{
             control_service_client::ControlServiceClient, watch_events_reply::Event,
-            ListGuestsRequest, WatchEventsRequest,
+            ListGuestsRequest,
         },
     },
 };
@@ -50,12 +50,11 @@ pub struct AutoNetworkChangeset {
 }
 
 impl AutoNetworkWatcher {
-    pub async fn new(mut control: ControlServiceClient<Channel>) -> Result<AutoNetworkWatcher> {
-        let watch_events_response = control.watch_events(WatchEventsRequest {}).await?;
-
+    pub async fn new(control: ControlServiceClient<Channel>) -> Result<AutoNetworkWatcher> {
+        let client = control.clone();
         Ok(AutoNetworkWatcher {
             control,
-            events: EventStream::open(watch_events_response.into_inner()).await?,
+            events: EventStream::open(client).await?,
             known: HashMap::new(),
         })
     }
@@ -136,7 +135,15 @@ impl AutoNetworkWatcher {
         let mut added: Vec<NetworkMetadata> = Vec::new();
         let mut removed: Vec<NetworkMetadata> = Vec::new();
 
-        for network in self.read().await? {
+        let networks = match self.read().await {
+            Ok(networks) => networks,
+            Err(error) => {
+                warn!("failed to read network changes: {}", error);
+                return Ok(AutoNetworkChangeset { added, removed });
+            }
+        };
+
+        for network in networks {
             seen.push(network.uuid);
             if self.known.contains_key(&network.uuid) {
                 continue;

crates/oci/Cargo.toml

@@ -2,7 +2,7 @@
 name = "krata-oci"
 description = "OCI services for the krata hypervisor."
 license.workspace = true
-version= "0.0.3"
+version.workspace = true
 homepage.workspace = true
 repository.workspace = true
 edition = "2021"
@@ -14,11 +14,13 @@ async-compression = { workspace = true, features = ["tokio", "gzip", "zstd"] }
 async-trait = { workspace = true }
 backhand = { workspace = true }
 bytes = { workspace = true }
+indexmap = { workspace = true }
 krata-tokio-tar = { workspace = true }
 log = { workspace = true }
 oci-spec = { workspace = true }
 path-clean = { workspace = true }
 reqwest = { workspace = true }
+scopeguard = { workspace = true }
 serde = { workspace = true }
 serde_json = { workspace = true }
 sha256 = { workspace = true }

crates/oci/examples/squashify.rs

@@ -2,7 +2,12 @@ use std::{env::args, path::PathBuf};
 
 use anyhow::Result;
 use env_logger::Env;
-use krataoci::{cache::ImageCache, compiler::ImageCompiler, name::ImageName};
+use krataoci::{
+    name::ImageName,
+    packer::{service::OciPackerService, OciPackedFormat},
+    progress::OciProgressContext,
+    registry::OciPlatform,
+};
 use tokio::fs;
 
 #[tokio::main]
@@ -17,13 +22,27 @@ async fn main() -> Result<()> {
         fs::create_dir(&cache_dir).await?;
     }
 
-    let cache = ImageCache::new(&cache_dir)?;
-    let compiler = ImageCompiler::new(&cache, seed)?;
-    let info = compiler.compile(&image).await?;
+    let (context, mut receiver) = OciProgressContext::create();
+    tokio::task::spawn(async move {
+        loop {
+            if (receiver.changed().await).is_err() {
+                break;
+            }
+            let progress = receiver.borrow_and_update();
+            println!("phase {:?}", progress.phase);
+            for (id, layer) in &progress.layers {
+                println!("{} {:?} {:?}", id, layer.phase, layer.indication,)
+            }
+        }
+    });
+    let service = OciPackerService::new(seed, &cache_dir, OciPlatform::current()).await?;
+    let packed = service
+        .request(image.clone(), OciPackedFormat::Squashfs, false, context)
+        .await?;
     println!(
         "generated squashfs of {} to {}",
         image,
-        info.image_squashfs.to_string_lossy()
+        packed.path.to_string_lossy()
     );
     Ok(())
 }

crates/oci/src/assemble.rs

@@ -0,0 +1,273 @@
+use crate::fetch::{OciImageFetcher, OciImageLayer, OciImageLayerReader, OciResolvedImage};
+use crate::progress::OciBoundProgress;
+use crate::schema::OciSchema;
+use crate::vfs::{VfsNode, VfsTree};
+use anyhow::{anyhow, Result};
+use log::{debug, trace, warn};
+use oci_spec::image::{Descriptor, ImageConfiguration, ImageManifest};
+
+use std::path::{Path, PathBuf};
+use std::pin::Pin;
+use std::sync::atomic::{AtomicBool, Ordering};
+use std::sync::Arc;
+use tokio::fs;
+use tokio_stream::StreamExt;
+use tokio_tar::{Archive, Entry};
+use uuid::Uuid;
+
+pub struct OciImageAssembled {
+    pub digest: String,
+    pub descriptor: Descriptor,
+    pub manifest: OciSchema<ImageManifest>,
+    pub config: OciSchema<ImageConfiguration>,
+    pub vfs: Arc<VfsTree>,
+    pub tmp_dir: Option<PathBuf>,
+}
+
+impl Drop for OciImageAssembled {
+    fn drop(&mut self) {
+        if let Some(tmp) = self.tmp_dir.clone() {
+            tokio::task::spawn(async move {
+                let _ = fs::remove_dir_all(&tmp).await;
+            });
+        }
+    }
+}
+
+pub struct OciImageAssembler {
+    downloader: OciImageFetcher,
+    resolved: Option<OciResolvedImage>,
+    progress: OciBoundProgress,
+    work_dir: PathBuf,
+    disk_dir: PathBuf,
+    tmp_dir: Option<PathBuf>,
+    success: AtomicBool,
+}
+
+impl OciImageAssembler {
+    pub async fn new(
+        downloader: OciImageFetcher,
+        resolved: OciResolvedImage,
+        progress: OciBoundProgress,
+        work_dir: Option<PathBuf>,
+        disk_dir: Option<PathBuf>,
+    ) -> Result<OciImageAssembler> {
+        let tmp_dir = if work_dir.is_none() || disk_dir.is_none() {
+            let mut tmp_dir = std::env::temp_dir().clone();
+            tmp_dir.push(format!("oci-assemble-{}", Uuid::new_v4()));
+            Some(tmp_dir)
+        } else {
+            None
+        };
+
+        let work_dir = if let Some(work_dir) = work_dir {
+            work_dir
+        } else {
+            let mut tmp_dir = tmp_dir
+                .clone()
+                .ok_or(anyhow!("tmp_dir was not created when expected"))?;
+            tmp_dir.push("work");
+            tmp_dir
+        };
+
+        let target_dir = if let Some(target_dir) = disk_dir {
+            target_dir
+        } else {
+            let mut tmp_dir = tmp_dir
+                .clone()
+                .ok_or(anyhow!("tmp_dir was not created when expected"))?;
+            tmp_dir.push("image");
+            tmp_dir
+        };
+
+        fs::create_dir_all(&work_dir).await?;
+        fs::create_dir_all(&target_dir).await?;
+
+        Ok(OciImageAssembler {
+            downloader,
+            resolved: Some(resolved),
+            progress,
+            work_dir,
+            disk_dir: target_dir,
+            tmp_dir,
+            success: AtomicBool::new(false),
+        })
+    }
+
+    pub async fn assemble(self) -> Result<OciImageAssembled> {
+        debug!("assemble");
+        let mut layer_dir = self.work_dir.clone();
+        layer_dir.push("layer");
+        fs::create_dir_all(&layer_dir).await?;
+        self.assemble_with(&layer_dir).await
+    }
+
+    async fn assemble_with(mut self, layer_dir: &Path) -> Result<OciImageAssembled> {
+        let Some(ref resolved) = self.resolved else {
+            return Err(anyhow!("resolved image was not available when expected"));
+        };
+        let local = self.downloader.download(resolved, layer_dir).await?;
+        let mut vfs = VfsTree::new();
+        for layer in &local.layers {
+            debug!(
+                "process layer digest={} compression={:?}",
+                &layer.digest, layer.compression,
+            );
+            self.progress
+                .update(|progress| {
+                    progress.start_extracting_layer(&layer.digest);
+                })
+                .await;
+            debug!("process layer digest={}", &layer.digest,);
+            let mut archive = layer.archive().await?;
+            let mut entries = archive.entries()?;
+            let mut count = 0u64;
+            let mut size = 0u64;
+            while let Some(entry) = entries.next().await {
+                let mut entry = entry?;
+                let path = entry.path()?;
+                let Some(name) = path.file_name() else {
+                    continue;
+                };
+                let Some(name) = name.to_str() else {
+                    continue;
+                };
+                if name.starts_with(".wh.") {
+                    self.process_whiteout_entry(&mut vfs, &entry, name, layer)
+                        .await?;
+                } else {
+                    let reference = vfs.insert_tar_entry(&entry)?;
+                    self.progress
+                        .update(|progress| {
+                            progress.extracting_layer(&layer.digest, &reference.name);
+                        })
+                        .await;
+                    size += self
+                        .process_write_entry(&mut vfs, &mut entry, layer)
+                        .await?;
+                    count += 1;
+                }
+            }
+            self.progress
+                .update(|progress| {
+                    progress.extracted_layer(&layer.digest, count, size);
+                })
+                .await;
+        }
+        for layer in &local.layers {
+            if layer.path.exists() {
+                fs::remove_file(&layer.path).await?;
+            }
+        }
+
+        let Some(resolved) = self.resolved.take() else {
+            return Err(anyhow!("resolved image was not available when expected"));
+        };
+
+        let assembled = OciImageAssembled {
+            vfs: Arc::new(vfs),
+            descriptor: resolved.descriptor,
+            digest: resolved.digest,
+            manifest: resolved.manifest,
+            config: local.config,
+            tmp_dir: self.tmp_dir.clone(),
+        };
+        self.success.store(true, Ordering::Release);
+        Ok(assembled)
+    }
+
+    async fn process_whiteout_entry(
+        &self,
+        vfs: &mut VfsTree,
+        entry: &Entry<Archive<Pin<Box<dyn OciImageLayerReader + Send>>>>,
+        name: &str,
+        layer: &OciImageLayer,
+    ) -> Result<()> {
+        let path = entry.path()?;
+        let mut path = path.to_path_buf();
+        path.pop();
+
+        let opaque = name == ".wh..wh..opq";
+
+        if !opaque {
+            let file = &name[4..];
+            path.push(file);
+        }
+
+        trace!(
+            "whiteout entry {:?} layer={} path={:?}",
+            entry.path()?,
+            &layer.digest,
+            path
+        );
+
+        let result = vfs.root.remove(&path);
+        if let Some((parent, mut removed)) = result {
+            delete_disk_paths(&removed).await?;
+            if opaque {
+                removed.children.clear();
+                parent.children.push(removed);
+            }
+        } else {
+            warn!(
+                "whiteout entry layer={} path={:?} did not exist",
+                &layer.digest, path
+            );
+        }
+        Ok(())
+    }
+
+    async fn process_write_entry(
+        &self,
+        vfs: &mut VfsTree,
+        entry: &mut Entry<Archive<Pin<Box<dyn OciImageLayerReader + Send>>>>,
+        layer: &OciImageLayer,
+    ) -> Result<u64> {
+        if !entry.header().entry_type().is_file() {
+            return Ok(0);
+        }
+        trace!(
+            "unpack entry layer={} path={:?} type={:?}",
+            &layer.digest,
+            entry.path()?,
+            entry.header().entry_type(),
+        );
+        entry.set_preserve_permissions(false);
+        entry.set_unpack_xattrs(false);
+        entry.set_preserve_mtime(false);
+        let path = entry
+            .unpack_in(&self.disk_dir)
+            .await?
+            .ok_or(anyhow!("unpack did not return a path"))?;
+        vfs.set_disk_path(&entry.path()?, &path)?;
+        Ok(entry.header().size()?)
+    }
+}
+
+impl Drop for OciImageAssembler {
+    fn drop(&mut self) {
+        if !self.success.load(Ordering::Acquire) {
+            if let Some(tmp_dir) = self.tmp_dir.clone() {
+                tokio::task::spawn(async move {
+                    let _ = fs::remove_dir_all(tmp_dir).await;
+                });
+            }
+        }
+    }
+}
+
+async fn delete_disk_paths(node: &VfsNode) -> Result<()> {
+    let mut queue = vec![node];
+    while !queue.is_empty() {
+        let node = queue.remove(0);
+        if let Some(ref disk_path) = node.disk_path {
+            if !disk_path.exists() {
+                warn!("disk path {:?} does not exist", disk_path);
+            }
+            fs::remove_file(disk_path).await?;
+        }
+        let children = node.children.iter().collect::<Vec<_>>();
+        queue.extend_from_slice(&children);
+    }
+    Ok(())
+}

crates/oci/src/cache.rs

@@ -1,70 +0,0 @@
-use super::compiler::ImageInfo;
-use anyhow::Result;
-use log::debug;
-use oci_spec::image::{ImageConfiguration, ImageManifest};
-use std::path::{Path, PathBuf};
-use tokio::fs;
-
-pub struct ImageCache {
-    cache_dir: PathBuf,
-}
-
-impl ImageCache {
-    pub fn new(cache_dir: &Path) -> Result<ImageCache> {
-        Ok(ImageCache {
-            cache_dir: cache_dir.to_path_buf(),
-        })
-    }
-
-    pub async fn recall(&self, digest: &str) -> Result<Option<ImageInfo>> {
-        let mut squashfs_path = self.cache_dir.clone();
-        let mut config_path = self.cache_dir.clone();
-        let mut manifest_path = self.cache_dir.clone();
-        squashfs_path.push(format!("{}.squashfs", digest));
-        manifest_path.push(format!("{}.manifest.json", digest));
-        config_path.push(format!("{}.config.json", digest));
-        Ok(
-            if squashfs_path.exists() && manifest_path.exists() && config_path.exists() {
-                let squashfs_metadata = fs::metadata(&squashfs_path).await?;
-                let manifest_metadata = fs::metadata(&manifest_path).await?;
-                let config_metadata = fs::metadata(&config_path).await?;
-                if squashfs_metadata.is_file()
-                    && manifest_metadata.is_file()
-                    && config_metadata.is_file()
-                {
-                    let manifest_text = fs::read_to_string(&manifest_path).await?;
-                    let manifest: ImageManifest = serde_json::from_str(&manifest_text)?;
-                    let config_text = fs::read_to_string(&config_path).await?;
-                    let config: ImageConfiguration = serde_json::from_str(&config_text)?;
-                    debug!("cache hit digest={}", digest);
-                    Some(ImageInfo::new(squashfs_path.clone(), manifest, config)?)
-                } else {
-                    None
-                }
-            } else {
-                debug!("cache miss digest={}", digest);
-                None
-            },
-        )
-    }
-
-    pub async fn store(&self, digest: &str, info: &ImageInfo) -> Result<ImageInfo> {
-        debug!("cache store digest={}", digest);
-        let mut squashfs_path = self.cache_dir.clone();
-        let mut manifest_path = self.cache_dir.clone();
-        let mut config_path = self.cache_dir.clone();
-        squashfs_path.push(format!("{}.squashfs", digest));
-        manifest_path.push(format!("{}.manifest.json", digest));
-        config_path.push(format!("{}.config.json", digest));
-        fs::copy(&info.image_squashfs, &squashfs_path).await?;
-        let manifest_text = serde_json::to_string_pretty(&info.manifest)?;
-        fs::write(&manifest_path, manifest_text).await?;
-        let config_text = serde_json::to_string_pretty(&info.config)?;
-        fs::write(&config_path, config_text).await?;
-        ImageInfo::new(
-            squashfs_path.clone(),
-            info.manifest.clone(),
-            info.config.clone(),
-        )
-    }
-}

crates/oci/src/compiler.rs

@@ -1,411 +0,0 @@
-use crate::cache::ImageCache;
-use crate::fetch::{OciImageDownloader, OciImageLayer};
-use crate::name::ImageName;
-use crate::registry::OciRegistryPlatform;
-use anyhow::{anyhow, Result};
-use backhand::compression::Compressor;
-use backhand::{FilesystemCompressor, FilesystemWriter, NodeHeader};
-use log::{debug, trace, warn};
-use oci_spec::image::{ImageConfiguration, ImageManifest};
-use std::borrow::Cow;
-use std::fs::File;
-use std::io::{BufWriter, ErrorKind, Read};
-use std::os::unix::fs::{FileTypeExt, MetadataExt, PermissionsExt};
-use std::path::{Path, PathBuf};
-use std::pin::Pin;
-use tokio::fs;
-use tokio::io::AsyncRead;
-use tokio_stream::StreamExt;
-use tokio_tar::{Archive, Entry};
-use uuid::Uuid;
-use walkdir::WalkDir;
-
-pub const IMAGE_SQUASHFS_VERSION: u64 = 2;
-
-pub struct ImageInfo {
-    pub image_squashfs: PathBuf,
-    pub manifest: ImageManifest,
-    pub config: ImageConfiguration,
-}
-
-impl ImageInfo {
-    pub fn new(
-        squashfs: PathBuf,
-        manifest: ImageManifest,
-        config: ImageConfiguration,
-    ) -> Result<ImageInfo> {
-        Ok(ImageInfo {
-            image_squashfs: squashfs,
-            manifest,
-            config,
-        })
-    }
-}
-
-pub struct ImageCompiler<'a> {
-    cache: &'a ImageCache,
-    seed: Option<PathBuf>,
-}
-
-impl ImageCompiler<'_> {
-    pub fn new(cache: &ImageCache, seed: Option<PathBuf>) -> Result<ImageCompiler> {
-        Ok(ImageCompiler { cache, seed })
-    }
-
-    pub async fn compile(&self, image: &ImageName) -> Result<ImageInfo> {
-        debug!("compile image={image}");
-        let mut tmp_dir = std::env::temp_dir().clone();
-        tmp_dir.push(format!("krata-compile-{}", Uuid::new_v4()));
-
-        let mut image_dir = tmp_dir.clone();
-        image_dir.push("image");
-        fs::create_dir_all(&image_dir).await?;
-
-        let mut layer_dir = tmp_dir.clone();
-        layer_dir.push("layer");
-        fs::create_dir_all(&layer_dir).await?;
-
-        let mut squash_file = tmp_dir.clone();
-        squash_file.push("image.squashfs");
-        let info = self
-            .download_and_compile(image, &layer_dir, &image_dir, &squash_file)
-            .await?;
-        fs::remove_dir_all(&tmp_dir).await?;
-        Ok(info)
-    }
-
-    async fn download_and_compile(
-        &self,
-        image: &ImageName,
-        layer_dir: &Path,
-        image_dir: &Path,
-        squash_file: &Path,
-    ) -> Result<ImageInfo> {
-        let downloader = OciImageDownloader::new(
-            self.seed.clone(),
-            layer_dir.to_path_buf(),
-            OciRegistryPlatform::current(),
-        );
-        let resolved = downloader.resolve(image.clone()).await?;
-        let cache_key = format!(
-            "manifest={}:squashfs-version={}\n",
-            resolved.digest, IMAGE_SQUASHFS_VERSION
-        );
-        let cache_digest = sha256::digest(cache_key);
-
-        if let Some(cached) = self.cache.recall(&cache_digest).await? {
-            return Ok(cached);
-        }
-
-        let local = downloader.download(resolved).await?;
-        for layer in &local.layers {
-            debug!(
-                "process layer digest={} compression={:?}",
-                &layer.digest, layer.compression,
-            );
-            let whiteouts = self.process_layer_whiteout(layer, image_dir).await?;
-            debug!(
-                "process layer digest={} whiteouts={:?}",
-                &layer.digest, whiteouts
-            );
-            let mut archive = layer.archive().await?;
-            let mut entries = archive.entries()?;
-            while let Some(entry) = entries.next().await {
-                let mut entry = entry?;
-                let path = entry.path()?;
-                let mut maybe_whiteout_path_str =
-                    path.to_str().map(|x| x.to_string()).unwrap_or_default();
-                if whiteouts.contains(&maybe_whiteout_path_str) {
-                    continue;
-                }
-                maybe_whiteout_path_str.push('/');
-                if whiteouts.contains(&maybe_whiteout_path_str) {
-                    continue;
-                }
-                let Some(name) = path.file_name() else {
-                    return Err(anyhow!("unable to get file name"));
-                };
-                let Some(name) = name.to_str() else {
-                    return Err(anyhow!("unable to get file name as string"));
-                };
-
-                if name.starts_with(".wh.") {
-                    continue;
-                } else {
-                    self.process_write_entry(&mut entry, layer, image_dir)
-                        .await?;
-                }
-            }
-        }
-
-        for layer in &local.layers {
-            if layer.path.exists() {
-                fs::remove_file(&layer.path).await?;
-            }
-        }
-
-        self.squash(image_dir, squash_file)?;
-        let info = ImageInfo::new(
-            squash_file.to_path_buf(),
-            local.image.manifest,
-            local.config,
-        )?;
-        self.cache.store(&cache_digest, &info).await
-    }
-
-    async fn process_layer_whiteout(
-        &self,
-        layer: &OciImageLayer,
-        image_dir: &Path,
-    ) -> Result<Vec<String>> {
-        let mut whiteouts = Vec::new();
-        let mut archive = layer.archive().await?;
-        let mut entries = archive.entries()?;
-        while let Some(entry) = entries.next().await {
-            let entry = entry?;
-            let path = entry.path()?;
-            let Some(name) = path.file_name() else {
-                return Err(anyhow!("unable to get file name"));
-            };
-            let Some(name) = name.to_str() else {
-                return Err(anyhow!("unable to get file name as string"));
-            };
-
-            if name.starts_with(".wh.") {
-                let path = self
-                    .process_whiteout_entry(&entry, name, layer, image_dir)
-                    .await?;
-                if let Some(path) = path {
-                    whiteouts.push(path);
-                }
-            }
-        }
-        Ok(whiteouts)
-    }
-
-    async fn process_whiteout_entry(
-        &self,
-        entry: &Entry<Archive<Pin<Box<dyn AsyncRead + Send>>>>,
-        name: &str,
-        layer: &OciImageLayer,
-        image_dir: &Path,
-    ) -> Result<Option<String>> {
-        let path = entry.path()?;
-        let mut dst = self.check_safe_entry(path.clone(), image_dir)?;
-        dst.pop();
-        let mut path = path.to_path_buf();
-        path.pop();
-
-        let opaque = name == ".wh..wh..opq";
-
-        if !opaque {
-            let file = &name[4..];
-            dst.push(file);
-            path.push(file);
-            self.check_safe_path(&dst, image_dir)?;
-        }
-
-        trace!("whiteout entry layer={} path={:?}", &layer.digest, path,);
-
-        let whiteout = path
-            .to_str()
-            .ok_or(anyhow!("unable to convert path to string"))?
-            .to_string();
-
-        if opaque {
-            if dst.is_dir() {
-                let mut reader = fs::read_dir(dst).await?;
-                while let Some(entry) = reader.next_entry().await? {
-                    let path = entry.path();
-                    if path.is_symlink() || path.is_file() {
-                        fs::remove_file(&path).await?;
-                    } else if path.is_dir() {
-                        fs::remove_dir_all(&path).await?;
-                    } else {
-                        return Err(anyhow!("opaque whiteout entry did not exist"));
-                    }
-                }
-            } else {
-                debug!(
-                    "whiteout opaque entry missing locally layer={} path={:?} local={:?}",
-                    &layer.digest,
-                    entry.path()?,
-                    dst,
-                );
-            }
-        } else if dst.is_file() || dst.is_symlink() {
-            fs::remove_file(&dst).await?;
-        } else if dst.is_dir() {
-            fs::remove_dir_all(&dst).await?;
-        } else {
-            debug!(
-                "whiteout entry missing locally layer={} path={:?} local={:?}",
-                &layer.digest,
-                entry.path()?,
-                dst,
-            );
-        }
-        Ok(if opaque { None } else { Some(whiteout) })
-    }
-
-    async fn process_write_entry(
-        &self,
-        entry: &mut Entry<Archive<Pin<Box<dyn AsyncRead + Send>>>>,
-        layer: &OciImageLayer,
-        image_dir: &Path,
-    ) -> Result<()> {
-        let uid = entry.header().uid()?;
-        let gid = entry.header().gid()?;
-        trace!(
-            "unpack entry layer={} path={:?} type={:?} uid={} gid={}",
-            &layer.digest,
-            entry.path()?,
-            entry.header().entry_type(),
-            uid,
-            gid,
-        );
-        entry.set_preserve_mtime(true);
-        entry.set_preserve_permissions(true);
-        entry.set_unpack_xattrs(true);
-        if let Some(path) = entry.unpack_in(image_dir).await? {
-            if !path.is_symlink() {
-                std::os::unix::fs::chown(path, Some(uid as u32), Some(gid as u32))?;
-            }
-        }
-        Ok(())
-    }
-
-    fn check_safe_entry(&self, path: Cow<Path>, image_dir: &Path) -> Result<PathBuf> {
-        let mut dst = image_dir.to_path_buf();
-        dst.push(path);
-        if let Some(name) = dst.file_name() {
-            if let Some(name) = name.to_str() {
-                if name.starts_with(".wh.") {
-                    let copy = dst.clone();
-                    dst.pop();
-                    self.check_safe_path(&dst, image_dir)?;
-                    return Ok(copy);
-                }
-            }
-        }
-        self.check_safe_path(&dst, image_dir)?;
-        Ok(dst)
-    }
-
-    fn check_safe_path(&self, dst: &Path, image_dir: &Path) -> Result<()> {
-        let resolved = path_clean::clean(dst);
-        if !resolved.starts_with(image_dir) {
-            return Err(anyhow!("layer attempts to work outside image dir"));
-        }
-        Ok(())
-    }
-
-    fn squash(&self, image_dir: &Path, squash_file: &Path) -> Result<()> {
-        let mut writer = FilesystemWriter::default();
-        writer.set_compressor(FilesystemCompressor::new(Compressor::Gzip, None)?);
-        let walk = WalkDir::new(image_dir).follow_links(false);
-        for entry in walk {
-            let entry = entry?;
-            let rel = entry
-                .path()
-                .strip_prefix(image_dir)?
-                .to_str()
-                .ok_or_else(|| anyhow!("failed to strip prefix of tmpdir"))?;
-            let rel = format!("/{}", rel);
-            trace!("squash write {}", rel);
-            let typ = entry.file_type();
-            let metadata = std::fs::symlink_metadata(entry.path())?;
-            let uid = metadata.uid();
-            let gid = metadata.gid();
-            let mode = metadata.permissions().mode();
-            let mtime = metadata.mtime();
-
-            if rel == "/" {
-                writer.set_root_uid(uid);
-                writer.set_root_gid(gid);
-                writer.set_root_mode(mode as u16);
-                continue;
-            }
-
-            let header = NodeHeader {
-                permissions: mode as u16,
-                uid,
-                gid,
-                mtime: mtime as u32,
-            };
-            if typ.is_symlink() {
-                let symlink = std::fs::read_link(entry.path())?;
-                let symlink = symlink
-                    .to_str()
-                    .ok_or_else(|| anyhow!("failed to read symlink"))?;
-                writer.push_symlink(symlink, rel, header)?;
-            } else if typ.is_dir() {
-                writer.push_dir(rel, header)?;
-            } else if typ.is_file() {
-                writer.push_file(ConsumingFileReader::new(entry.path()), rel, header)?;
-            } else if typ.is_block_device() {
-                let device = metadata.dev();
-                writer.push_block_device(device as u32, rel, header)?;
-            } else if typ.is_char_device() {
-                let device = metadata.dev();
-                writer.push_char_device(device as u32, rel, header)?;
-            } else if typ.is_fifo() {
-                writer.push_fifo(rel, header)?;
-            } else if typ.is_socket() {
-                writer.push_socket(rel, header)?;
-            } else {
-                return Err(anyhow!("invalid file type"));
-            }
-        }
-
-        let squash_file_path = squash_file
-            .to_str()
-            .ok_or_else(|| anyhow!("failed to convert squashfs string"))?;
-
-        let file = File::create(squash_file)?;
-        let mut bufwrite = BufWriter::new(file);
-        trace!("squash generate: {}", squash_file_path);
-        writer.write(&mut bufwrite)?;
-        std::fs::remove_dir_all(image_dir)?;
-        Ok(())
-    }
-}
-
-struct ConsumingFileReader {
-    path: PathBuf,
-    file: Option<File>,
-}
-
-impl ConsumingFileReader {
-    fn new(path: &Path) -> ConsumingFileReader {
-        ConsumingFileReader {
-            path: path.to_path_buf(),
-            file: None,
-        }
-    }
-}
-
-impl Read for ConsumingFileReader {
-    fn read(&mut self, buf: &mut [u8]) -> std::io::Result<usize> {
-        if self.file.is_none() {
-            self.file = Some(File::open(&self.path)?);
-        }
-        let Some(ref mut file) = self.file else {
-            return Err(std::io::Error::new(
-                ErrorKind::NotFound,
-                "file was not opened",
-            ));
-        };
-        file.read(buf)
-    }
-}
-
-impl Drop for ConsumingFileReader {
-    fn drop(&mut self) {
-        let file = self.file.take();
-        drop(file);
-        if let Err(error) = std::fs::remove_file(&self.path) {
-            warn!("failed to delete consuming file {:?}: {}", self.path, error);
-        }
-    }
-}

crates/oci/src/fetch.rs

@@ -1,9 +1,17 @@
+use crate::{
+    progress::{OciBoundProgress, OciProgressPhase},
+    schema::OciSchema,
+};
+
 use super::{
     name::ImageName,
-    registry::{OciRegistryClient, OciRegistryPlatform},
+    registry::{OciPlatform, OciRegistryClient},
 };
 
 use std::{
+    fmt::Debug,
+    io::SeekFrom,
+    os::unix::fs::MetadataExt,
     path::{Path, PathBuf},
     pin::Pin,
 };
@@ -12,20 +20,21 @@ use anyhow::{anyhow, Result};
 use async_compression::tokio::bufread::{GzipDecoder, ZstdDecoder};
 use log::debug;
 use oci_spec::image::{
-    Descriptor, ImageConfiguration, ImageIndex, ImageManifest, MediaType, ToDockerV2S2,
+    Descriptor, DescriptorBuilder, ImageConfiguration, ImageIndex, ImageManifest, MediaType,
+    ToDockerV2S2,
 };
 use serde::de::DeserializeOwned;
 use tokio::{
-    fs::File,
-    io::{AsyncRead, AsyncReadExt, BufReader, BufWriter},
+    fs::{self, File},
+    io::{AsyncRead, AsyncReadExt, AsyncSeekExt, BufReader, BufWriter},
 };
 use tokio_stream::StreamExt;
 use tokio_tar::Archive;
 
-pub struct OciImageDownloader {
+pub struct OciImageFetcher {
     seed: Option<PathBuf>,
-    storage: PathBuf,
-    platform: OciRegistryPlatform,
+    platform: OciPlatform,
+    progress: OciBoundProgress,
 }
 
 #[derive(Clone, Debug, PartialEq, Eq)]
@@ -37,16 +46,43 @@ pub enum OciImageLayerCompression {
 
 #[derive(Clone, Debug)]
 pub struct OciImageLayer {
+    pub metadata: Descriptor,
     pub path: PathBuf,
     pub digest: String,
     pub compression: OciImageLayerCompression,
 }
 
+#[async_trait::async_trait]
+pub trait OciImageLayerReader: AsyncRead + Sync {
+    async fn position(&mut self) -> Result<u64>;
+}
+
+#[async_trait::async_trait]
+impl OciImageLayerReader for BufReader<File> {
+    async fn position(&mut self) -> Result<u64> {
+        Ok(self.seek(SeekFrom::Current(0)).await?)
+    }
+}
+
+#[async_trait::async_trait]
+impl OciImageLayerReader for GzipDecoder<BufReader<File>> {
+    async fn position(&mut self) -> Result<u64> {
+        self.get_mut().position().await
+    }
+}
+
+#[async_trait::async_trait]
+impl OciImageLayerReader for ZstdDecoder<BufReader<File>> {
+    async fn position(&mut self) -> Result<u64> {
+        self.get_mut().position().await
+    }
+}
+
 impl OciImageLayer {
-    pub async fn decompress(&self) -> Result<Pin<Box<dyn AsyncRead + Send>>> {
+    pub async fn decompress(&self) -> Result<Pin<Box<dyn OciImageLayerReader + Send>>> {
         let file = File::open(&self.path).await?;
         let reader = BufReader::new(file);
-        let reader: Pin<Box<dyn AsyncRead + Send>> = match self.compression {
+        let reader: Pin<Box<dyn OciImageLayerReader + Send>> = match self.compression {
             OciImageLayerCompression::None => Box::pin(reader),
             OciImageLayerCompression::Gzip => Box::pin(GzipDecoder::new(reader)),
             OciImageLayerCompression::Zstd => Box::pin(ZstdDecoder::new(reader)),
@@ -54,7 +90,7 @@ impl OciImageLayer {
         Ok(reader)
     }
 
-    pub async fn archive(&self) -> Result<Archive<Pin<Box<dyn AsyncRead + Send>>>> {
+    pub async fn archive(&self) -> Result<Archive<Pin<Box<dyn OciImageLayerReader + Send>>>> {
         let decompress = self.decompress().await?;
         Ok(Archive::new(decompress))
     }
@@ -64,33 +100,34 @@ impl OciImageLayer {
 pub struct OciResolvedImage {
     pub name: ImageName,
     pub digest: String,
-    pub manifest: ImageManifest,
+    pub descriptor: Descriptor,
+    pub manifest: OciSchema<ImageManifest>,
 }
 
 #[derive(Clone, Debug)]
 pub struct OciLocalImage {
     pub image: OciResolvedImage,
-    pub config: ImageConfiguration,
+    pub config: OciSchema<ImageConfiguration>,
     pub layers: Vec<OciImageLayer>,
 }
 
-impl OciImageDownloader {
+impl OciImageFetcher {
     pub fn new(
         seed: Option<PathBuf>,
-        storage: PathBuf,
-        platform: OciRegistryPlatform,
-    ) -> OciImageDownloader {
-        OciImageDownloader {
+        platform: OciPlatform,
+        progress: OciBoundProgress,
+    ) -> OciImageFetcher {
+        OciImageFetcher {
             seed,
-            storage,
             platform,
+            progress,
         }
     }
 
-    async fn load_seed_json_blob<T: DeserializeOwned>(
+    async fn load_seed_json_blob<T: Clone + Debug + DeserializeOwned>(
         &self,
         descriptor: &Descriptor,
-    ) -> Result<Option<T>> {
+    ) -> Result<Option<OciSchema<T>>> {
         let digest = descriptor.digest();
         let Some((digest_type, digest_content)) = digest.split_once(':') else {
             return Err(anyhow!("digest content was not properly formatted"));
@@ -99,7 +136,10 @@ impl OciImageDownloader {
         self.load_seed_json(&want).await
     }
 
-    async fn load_seed_json<T: DeserializeOwned>(&self, want: &str) -> Result<Option<T>> {
+    async fn load_seed_json<T: Clone + Debug + DeserializeOwned>(
+        &self,
+        want: &str,
+    ) -> Result<Option<OciSchema<T>>> {
         let Some(ref seed) = self.seed else {
             return Ok(None);
         };
@@ -111,10 +151,10 @@ impl OciImageDownloader {
             let mut entry = entry?;
             let path = String::from_utf8(entry.path_bytes().to_vec())?;
             if path == want {
-                let mut content = String::new();
-                entry.read_to_string(&mut content).await?;
-                let data = serde_json::from_str::<T>(&content)?;
-                return Ok(Some(data));
+                let mut content = Vec::new();
+                entry.read_to_end(&mut content).await?;
+                let item = serde_json::from_slice::<T>(&content)?;
+                return Ok(Some(OciSchema::new(content, item)));
             }
         }
         Ok(None)
@@ -152,7 +192,7 @@ impl OciImageDownloader {
 
         if let Some(index) = self.load_seed_json::<ImageIndex>("index.json").await? {
             let mut found: Option<&Descriptor> = None;
-            for manifest in index.manifests() {
+            for manifest in index.item().manifests() {
                 let Some(annotations) = manifest.annotations() else {
                     continue;
                 };
@@ -177,6 +217,13 @@ impl OciImageDownloader {
                         continue;
                     }
                 }
+
+                if let Some(ref digest) = image.digest {
+                    if digest != manifest.digest() {
+                        continue;
+                    }
+                }
+
                 found = Some(manifest);
                 break;
             }
@@ -190,6 +237,7 @@ impl OciImageDownloader {
                     );
                     return Ok(OciResolvedImage {
                         name: image,
+                        descriptor: found.clone(),
                         digest: found.digest().clone(),
                         manifest,
                     });
@@ -198,37 +246,79 @@ impl OciImageDownloader {
         }
 
         let mut client = OciRegistryClient::new(image.registry_url()?, self.platform.clone())?;
-        let (manifest, digest) = client
-            .get_manifest_with_digest(&image.name, &image.reference)
+        let (manifest, descriptor, digest) = client
+            .get_manifest_with_digest(&image.name, image.reference.as_ref(), image.digest.as_ref())
             .await?;
+        let descriptor = descriptor.unwrap_or_else(|| {
+            DescriptorBuilder::default()
+                .media_type(MediaType::ImageManifest)
+                .size(manifest.raw().len() as i64)
+                .digest(digest.clone())
+                .build()
+                .unwrap()
+        });
         Ok(OciResolvedImage {
             name: image,
+            descriptor,
             digest,
             manifest,
         })
     }
 
-    pub async fn download(&self, image: OciResolvedImage) -> Result<OciLocalImage> {
-        let config: ImageConfiguration;
-
+    pub async fn download(
+        &self,
+        image: &OciResolvedImage,
+        layer_dir: &Path,
+    ) -> Result<OciLocalImage> {
+        let config: OciSchema<ImageConfiguration>;
+        self.progress
+            .update(|progress| {
+                progress.phase = OciProgressPhase::ConfigDownload;
+            })
+            .await;
         let mut client = OciRegistryClient::new(image.name.registry_url()?, self.platform.clone())?;
         if let Some(seeded) = self
-            .load_seed_json_blob::<ImageConfiguration>(image.manifest.config())
+            .load_seed_json_blob::<ImageConfiguration>(image.manifest.item().config())
             .await?
         {
             config = seeded;
         } else {
             let config_bytes = client
-                .get_blob(&image.name.name, image.manifest.config())
+                .get_blob(&image.name.name, image.manifest.item().config())
                 .await?;
-            config = serde_json::from_slice(&config_bytes)?;
+            config = OciSchema::new(
+                config_bytes.to_vec(),
+                serde_json::from_slice(&config_bytes)?,
+            );
         }
+        self.progress
+            .update(|progress| {
+                progress.phase = OciProgressPhase::LayerDownload;
+
+                for layer in image.manifest.item().layers() {
+                    progress.add_layer(layer.digest());
+                }
+            })
+            .await;
         let mut layers = Vec::new();
-        for layer in image.manifest.layers() {
-            layers.push(self.acquire_layer(&image.name, layer, &mut client).await?);
+        for layer in image.manifest.item().layers() {
+            self.progress
+                .update(|progress| {
+                    progress.downloading_layer(layer.digest(), 0, layer.size() as u64);
+                })
+                .await;
+            layers.push(
+                self.acquire_layer(&image.name, layer, layer_dir, &mut client)
+                    .await?,
+            );
+            self.progress
+                .update(|progress| {
+                    progress.downloaded_layer(layer.digest(), layer.size() as u64);
+                })
+                .await;
         }
         Ok(OciLocalImage {
-            image,
+            image: image.clone(),
             config,
             layers,
         })
@@ -238,6 +328,7 @@ impl OciImageDownloader {
         &self,
         image: &ImageName,
         layer: &Descriptor,
+        layer_dir: &Path,
         client: &mut OciRegistryClient,
     ) -> Result<OciImageLayer> {
         debug!(
@@ -245,13 +336,15 @@ impl OciImageDownloader {
             layer.digest(),
             layer.size()
         );
-        let mut layer_path = self.storage.clone();
+        let mut layer_path = layer_dir.to_path_buf();
         layer_path.push(format!("{}.layer", layer.digest()));
 
         let seeded = self.extract_seed_blob(layer, &layer_path).await?;
         if !seeded {
             let file = File::create(&layer_path).await?;
-            let size = client.write_blob_to_file(&image.name, layer, file).await?;
+            let size = client
+                .write_blob_to_file(&image.name, layer, file, Some(self.progress.clone()))
+                .await?;
             if layer.size() as u64 != size {
                 return Err(anyhow!(
                     "downloaded layer size differs from size in manifest",
@@ -259,6 +352,12 @@ impl OciImageDownloader {
             }
         }
 
+        let metadata = fs::metadata(&layer_path).await?;
+
+        if layer.size() as u64 != metadata.size() {
+            return Err(anyhow!("layer size differs from size in manifest",));
+        }
+
         let mut media_type = layer.media_type().clone();
 
         // docker layer compatibility
@@ -273,6 +372,7 @@ impl OciImageDownloader {
             other => return Err(anyhow!("found layer with unknown media type: {}", other)),
         };
         Ok(OciImageLayer {
+            metadata: layer.clone(),
             path: layer_path,
             digest: layer.digest().clone(),
             compression,

crates/oci/src/lib.rs

@@ -1,5 +1,8 @@
-pub mod cache;
-pub mod compiler;
+pub mod assemble;
 pub mod fetch;
 pub mod name;
+pub mod packer;
+pub mod progress;
 pub mod registry;
+pub mod schema;
+pub mod vfs;

crates/oci/src/name.rs

@@ -2,27 +2,39 @@ use anyhow::Result;
 use std::fmt;
 use url::Url;
 
-const DOCKER_HUB_MIRROR: &str = "mirror.gcr.io";
-const DEFAULT_IMAGE_TAG: &str = "latest";
-
 #[derive(Debug, Clone, PartialEq, Eq, Hash)]
 pub struct ImageName {
     pub hostname: String,
     pub port: Option<u16>,
     pub name: String,
-    pub reference: String,
+    pub reference: Option<String>,
+    pub digest: Option<String>,
 }
 
 impl fmt::Display for ImageName {
     fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
-        if let Some(port) = self.port {
-            write!(
-                f,
-                "{}:{}/{}:{}",
-                self.hostname, port, self.name, self.reference
-            )
+        let mut suffix = String::new();
+
+        if let Some(ref reference) = self.reference {
+            suffix.push(':');
+            suffix.push_str(reference);
+        }
+
+        if let Some(ref digest) = self.digest {
+            suffix.push('@');
+            suffix.push_str(digest);
+        }
+
+        if ImageName::DOCKER_HUB_MIRROR == self.hostname && self.port.is_none() {
+            if self.name.starts_with("library/") {
+                write!(f, "{}{}", &self.name[8..], suffix)
+            } else {
+                write!(f, "{}{}", self.name, suffix)
+            }
+        } else if let Some(port) = self.port {
+            write!(f, "{}:{}/{}{}", self.hostname, port, self.name, suffix)
         } else {
-            write!(f, "{}/{}:{}", self.hostname, self.name, self.reference)
+            write!(f, "{}/{}{}", self.hostname, self.name, suffix)
         }
     }
 }
@@ -35,13 +47,21 @@ impl Default for ImageName {
 }
 
 impl ImageName {
+    pub const DOCKER_HUB_MIRROR: &'static str = "registry.docker.io";
+    pub const DEFAULT_IMAGE_TAG: &'static str = "latest";
+
     pub fn parse(name: &str) -> Result<Self> {
         let full_name = name.to_string();
         let name = full_name.clone();
         let (mut hostname, mut name) = name
             .split_once('/')
             .map(|x| (x.0.to_string(), x.1.to_string()))
-            .unwrap_or_else(|| (DOCKER_HUB_MIRROR.to_string(), format!("library/{}", name)));
+            .unwrap_or_else(|| {
+                (
+                    ImageName::DOCKER_HUB_MIRROR.to_string(),
+                    format!("library/{}", name),
+                )
+            });
 
         // heuristic to find any docker hub image formats
         // that may be in the hostname format. for example:
@@ -49,7 +69,7 @@ impl ImageName {
         // and neither will abc/hello/xyz:latest
         if !hostname.contains('.') && full_name.chars().filter(|x| *x == '/').count() == 1 {
             name = format!("{}/{}", hostname, name);
-            hostname = DOCKER_HUB_MIRROR.to_string();
+            hostname = ImageName::DOCKER_HUB_MIRROR.to_string();
         }
 
         let (hostname, port) = if let Some((hostname, port)) = hostname
@@ -60,15 +80,54 @@ impl ImageName {
         } else {
             (hostname, None)
         };
-        let (name, reference) = name
-            .split_once(':')
-            .map(|x| (x.0.to_string(), x.1.to_string()))
-            .unwrap_or((name.to_string(), DEFAULT_IMAGE_TAG.to_string()));
+
+        let name_has_digest = if name.contains('@') {
+            let digest_start = name.chars().position(|c| c == '@');
+            let ref_start = name.chars().position(|c| c == ':');
+            if let (Some(digest_start), Some(ref_start)) = (digest_start, ref_start) {
+                digest_start < ref_start
+            } else {
+                true
+            }
+        } else {
+            false
+        };
+
+        let (name, digest) = if name_has_digest {
+            name.split_once('@')
+                .map(|(name, digest)| (name.to_string(), Some(digest.to_string())))
+                .unwrap_or_else(|| (name, None))
+        } else {
+            (name, None)
+        };
+
+        let (name, reference) = if name.contains(':') {
+            name.split_once(':')
+                .map(|(name, reference)| (name.to_string(), Some(reference.to_string())))
+                .unwrap_or((name, None))
+        } else {
+            (name, None)
+        };
+
+        let (reference, digest) = if let Some(reference) = reference {
+            if let Some(digest) = digest {
+                (Some(reference), Some(digest))
+            } else {
+                reference
+                    .split_once('@')
+                    .map(|(reff, digest)| (Some(reff.to_string()), Some(digest.to_string())))
+                    .unwrap_or_else(|| (Some(reference), None))
+            }
+        } else {
+            (None, digest)
+        };
+
         Ok(ImageName {
             hostname,
             port,
             name,
             reference,
+            digest,
         })
     }
 

crates/oci/src/packer/backend.rs

@@ -0,0 +1,236 @@
+use std::{os::unix::fs::MetadataExt, path::Path, process::Stdio, sync::Arc};
+
+use super::OciPackedFormat;
+use crate::{progress::OciBoundProgress, vfs::VfsTree};
+use anyhow::{anyhow, Result};
+use log::warn;
+use tokio::{
+    fs::{self, File},
+    io::BufWriter,
+    pin,
+    process::{Child, Command},
+    select,
+};
+
+#[derive(Debug, Clone, Copy)]
+pub enum OciPackerBackendType {
+    MkSquashfs,
+    MkfsErofs,
+    Tar,
+}
+
+impl OciPackerBackendType {
+    pub fn format(&self) -> OciPackedFormat {
+        match self {
+            OciPackerBackendType::MkSquashfs => OciPackedFormat::Squashfs,
+            OciPackerBackendType::MkfsErofs => OciPackedFormat::Erofs,
+            OciPackerBackendType::Tar => OciPackedFormat::Tar,
+        }
+    }
+
+    pub fn create(&self) -> Box<dyn OciPackerBackend> {
+        match self {
+            OciPackerBackendType::MkSquashfs => {
+                Box::new(OciPackerMkSquashfs {}) as Box<dyn OciPackerBackend>
+            }
+            OciPackerBackendType::MkfsErofs => {
+                Box::new(OciPackerMkfsErofs {}) as Box<dyn OciPackerBackend>
+            }
+            OciPackerBackendType::Tar => Box::new(OciPackerTar {}) as Box<dyn OciPackerBackend>,
+        }
+    }
+}
+
+#[async_trait::async_trait]
+pub trait OciPackerBackend: Send + Sync {
+    async fn pack(&self, progress: OciBoundProgress, vfs: Arc<VfsTree>, file: &Path) -> Result<()>;
+}
+
+pub struct OciPackerMkSquashfs {}
+
+#[async_trait::async_trait]
+impl OciPackerBackend for OciPackerMkSquashfs {
+    async fn pack(&self, progress: OciBoundProgress, vfs: Arc<VfsTree>, file: &Path) -> Result<()> {
+        progress
+            .update(|progress| {
+                progress.start_packing();
+            })
+            .await;
+
+        let child = Command::new("mksquashfs")
+            .arg("-")
+            .arg(file)
+            .arg("-comp")
+            .arg("gzip")
+            .arg("-tar")
+            .stdin(Stdio::piped())
+            .stderr(Stdio::null())
+            .stdout(Stdio::null())
+            .spawn()?;
+        let mut child = ChildProcessKillGuard(child);
+        let stdin = child
+            .0
+            .stdin
+            .take()
+            .ok_or(anyhow!("unable to acquire stdin stream"))?;
+        let mut writer = Some(tokio::task::spawn(async move {
+            if let Err(error) = vfs.write_to_tar(stdin).await {
+                warn!("failed to write tar: {}", error);
+                return Err(error);
+            }
+            Ok(())
+        }));
+        let wait = child.0.wait();
+        pin!(wait);
+        let status_result = loop {
+            if let Some(inner) = writer.as_mut() {
+                select! {
+                    x = inner => {
+                        writer = None;
+                        match x {
+                            Ok(_) => {},
+                            Err(error) => {
+                                return Err(error.into());
+                            }
+                        }
+                    },
+                    status = &mut wait => {
+                        break status;
+                    }
+                };
+            } else {
+                select! {
+                    status = &mut wait => {
+                        break status;
+                    }
+                };
+            }
+        };
+        if let Some(writer) = writer {
+            writer.await??;
+        }
+        let status = status_result?;
+        if !status.success() {
+            Err(anyhow!(
+                "mksquashfs failed with exit code: {}",
+                status.code().unwrap()
+            ))
+        } else {
+            let metadata = fs::metadata(&file).await?;
+            progress
+                .update(|progress| progress.complete(metadata.size()))
+                .await;
+            Ok(())
+        }
+    }
+}
+
+pub struct OciPackerMkfsErofs {}
+
+#[async_trait::async_trait]
+impl OciPackerBackend for OciPackerMkfsErofs {
+    async fn pack(&self, progress: OciBoundProgress, vfs: Arc<VfsTree>, file: &Path) -> Result<()> {
+        progress
+            .update(|progress| {
+                progress.start_packing();
+            })
+            .await;
+
+        let child = Command::new("mkfs.erofs")
+            .arg("-L")
+            .arg("root")
+            .arg("--tar=-")
+            .arg(file)
+            .stdin(Stdio::piped())
+            .stderr(Stdio::null())
+            .stdout(Stdio::null())
+            .spawn()?;
+        let mut child = ChildProcessKillGuard(child);
+        let stdin = child
+            .0
+            .stdin
+            .take()
+            .ok_or(anyhow!("unable to acquire stdin stream"))?;
+        let mut writer = Some(tokio::task::spawn(
+            async move { vfs.write_to_tar(stdin).await },
+        ));
+        let wait = child.0.wait();
+        pin!(wait);
+        let status_result = loop {
+            if let Some(inner) = writer.as_mut() {
+                select! {
+                    x = inner => {
+                        match x {
+                            Ok(_) => {
+                                writer = None;
+                            },
+                            Err(error) => {
+                                return Err(error.into());
+                            }
+                        }
+                    },
+                    status = &mut wait => {
+                        break status;
+                    }
+                };
+            } else {
+                select! {
+                    status = &mut wait => {
+                        break status;
+                    }
+                };
+            }
+        };
+        if let Some(writer) = writer {
+            writer.await??;
+        }
+        let status = status_result?;
+        if !status.success() {
+            Err(anyhow!(
+                "mkfs.erofs failed with exit code: {}",
+                status.code().unwrap()
+            ))
+        } else {
+            let metadata = fs::metadata(&file).await?;
+            progress
+                .update(|progress| {
+                    progress.complete(metadata.size());
+                })
+                .await;
+            Ok(())
+        }
+    }
+}
+
+pub struct OciPackerTar {}
+
+#[async_trait::async_trait]
+impl OciPackerBackend for OciPackerTar {
+    async fn pack(&self, progress: OciBoundProgress, vfs: Arc<VfsTree>, file: &Path) -> Result<()> {
+        progress
+            .update(|progress| {
+                progress.start_packing();
+            })
+            .await;
+
+        let output = File::create(file).await?;
+        let output = BufWriter::new(output);
+        vfs.write_to_tar(output).await?;
+
+        let metadata = fs::metadata(file).await?;
+        progress
+            .update(|progress| {
+                progress.complete(metadata.size());
+            })
+            .await;
+        Ok(())
+    }
+}
+
+struct ChildProcessKillGuard(Child);
+
+impl Drop for ChildProcessKillGuard {
+    fn drop(&mut self) {
+        let _ = self.0.start_kill();
+    }
+}

crates/oci/src/packer/cache.rs

@@ -0,0 +1,220 @@
+use crate::{
+    name::ImageName,
+    packer::{OciPackedFormat, OciPackedImage},
+    schema::OciSchema,
+};
+
+use anyhow::Result;
+use log::{debug, error};
+use oci_spec::image::{
+    Descriptor, ImageConfiguration, ImageIndex, ImageIndexBuilder, ImageManifest, MediaType,
+    ANNOTATION_REF_NAME,
+};
+use std::{
+    path::{Path, PathBuf},
+    sync::Arc,
+};
+use tokio::{fs, sync::RwLock};
+
+#[derive(Clone)]
+pub struct OciPackerCache {
+    cache_dir: PathBuf,
+    index: Arc<RwLock<ImageIndex>>,
+}
+
+const ANNOTATION_IMAGE_NAME: &str = "io.containerd.image.name";
+const ANNOTATION_OCI_PACKER_FORMAT: &str = "dev.krata.oci.packer.format";
+
+impl OciPackerCache {
+    pub async fn new(cache_dir: &Path) -> Result<OciPackerCache> {
+        let index = ImageIndexBuilder::default()
+            .schema_version(2u32)
+            .media_type(MediaType::ImageIndex)
+            .manifests(Vec::new())
+            .build()?;
+        let cache = OciPackerCache {
+            cache_dir: cache_dir.to_path_buf(),
+            index: Arc::new(RwLock::new(index)),
+        };
+
+        {
+            let mut mutex = cache.index.write().await;
+            *mutex = cache.load_index().await?;
+        }
+
+        Ok(cache)
+    }
+
+    pub async fn list(&self) -> Result<Vec<Descriptor>> {
+        let index = self.index.read().await;
+        Ok(index.manifests().clone())
+    }
+
+    pub async fn recall(
+        &self,
+        name: ImageName,
+        digest: &str,
+        format: OciPackedFormat,
+    ) -> Result<Option<OciPackedImage>> {
+        let index = self.index.read().await;
+
+        let mut descriptor: Option<Descriptor> = None;
+        for manifest in index.manifests() {
+            if manifest.digest() == digest
+                && manifest
+                    .annotations()
+                    .as_ref()
+                    .and_then(|x| x.get(ANNOTATION_OCI_PACKER_FORMAT))
+                    .map(|x| x.as_str())
+                    == Some(format.extension())
+            {
+                descriptor = Some(manifest.clone());
+                break;
+            }
+        }
+
+        let Some(descriptor) = descriptor else {
+            return Ok(None);
+        };
+
+        let mut fs_path = self.cache_dir.clone();
+        let mut config_path = self.cache_dir.clone();
+        let mut manifest_path = self.cache_dir.clone();
+        fs_path.push(format!("{}.{}", digest, format.extension()));
+        manifest_path.push(format!("{}.manifest.json", digest));
+        config_path.push(format!("{}.config.json", digest));
+
+        if fs_path.exists() && manifest_path.exists() && config_path.exists() {
+            let image_metadata = fs::metadata(&fs_path).await?;
+            let manifest_metadata = fs::metadata(&manifest_path).await?;
+            let config_metadata = fs::metadata(&config_path).await?;
+            if image_metadata.is_file() && manifest_metadata.is_file() && config_metadata.is_file()
+            {
+                let manifest_bytes = fs::read(&manifest_path).await?;
+                let manifest: ImageManifest = serde_json::from_slice(&manifest_bytes)?;
+                let config_bytes = fs::read(&config_path).await?;
+                let config: ImageConfiguration = serde_json::from_slice(&config_bytes)?;
+                debug!("cache hit digest={}", digest);
+                Ok(Some(OciPackedImage::new(
+                    name,
+                    digest.to_string(),
+                    fs_path.clone(),
+                    format,
+                    descriptor,
+                    OciSchema::new(config_bytes, config),
+                    OciSchema::new(manifest_bytes, manifest),
+                )))
+            } else {
+                Ok(None)
+            }
+        } else {
+            debug!("cache miss digest={}", digest);
+            Ok(None)
+        }
+    }
+
+    pub async fn store(&self, packed: OciPackedImage) -> Result<OciPackedImage> {
+        let mut index = self.index.write().await;
+        let mut manifests = index.manifests().clone();
+        debug!("cache store digest={}", packed.digest);
+        let mut fs_path = self.cache_dir.clone();
+        let mut manifest_path = self.cache_dir.clone();
+        let mut config_path = self.cache_dir.clone();
+        fs_path.push(format!("{}.{}", packed.digest, packed.format.extension()));
+        manifest_path.push(format!("{}.manifest.json", packed.digest));
+        config_path.push(format!("{}.config.json", packed.digest));
+        if fs::rename(&packed.path, &fs_path).await.is_err() {
+            fs::copy(&packed.path, &fs_path).await?;
+            fs::remove_file(&packed.path).await?;
+        }
+        fs::write(&config_path, packed.config.raw()).await?;
+        fs::write(&manifest_path, packed.manifest.raw()).await?;
+        manifests.retain(|item| {
+            if item.digest() != &packed.digest {
+                return true;
+            }
+
+            let Some(format) = item
+                .annotations()
+                .as_ref()
+                .and_then(|x| x.get(ANNOTATION_OCI_PACKER_FORMAT))
+                .map(|x| x.as_str())
+            else {
+                return true;
+            };
+
+            if format != packed.format.extension() {
+                return true;
+            }
+
+            false
+        });
+
+        let mut descriptor = packed.descriptor.clone();
+        let mut annotations = descriptor.annotations().clone().unwrap_or_default();
+        annotations.insert(
+            ANNOTATION_OCI_PACKER_FORMAT.to_string(),
+            packed.format.extension().to_string(),
+        );
+        let image_name = packed.name.to_string();
+        annotations.insert(ANNOTATION_IMAGE_NAME.to_string(), image_name);
+        let image_ref = packed.name.reference.clone();
+        if let Some(image_ref) = image_ref {
+            annotations.insert(ANNOTATION_REF_NAME.to_string(), image_ref);
+        }
+        descriptor.set_annotations(Some(annotations));
+        manifests.push(descriptor.clone());
+        index.set_manifests(manifests);
+        self.save_index(&index).await?;
+
+        let packed = OciPackedImage::new(
+            packed.name,
+            packed.digest,
+            fs_path.clone(),
+            packed.format,
+            descriptor,
+            packed.config,
+            packed.manifest,
+        );
+        Ok(packed)
+    }
+
+    async fn save_empty_index(&self) -> Result<ImageIndex> {
+        let index = ImageIndexBuilder::default()
+            .schema_version(2u32)
+            .media_type(MediaType::ImageIndex)
+            .manifests(Vec::new())
+            .build()?;
+        self.save_index(&index).await?;
+        Ok(index)
+    }
+
+    async fn load_index(&self) -> Result<ImageIndex> {
+        let mut index_path = self.cache_dir.clone();
+        index_path.push("index.json");
+
+        if !index_path.exists() {
+            self.save_empty_index().await?;
+        }
+
+        let content = fs::read_to_string(&index_path).await?;
+        let index = match serde_json::from_str::<ImageIndex>(&content) {
+            Ok(index) => index,
+            Err(error) => {
+                error!("image index was corrupted, creating a new one: {}", error);
+                self.save_empty_index().await?
+            }
+        };
+
+        Ok(index)
+    }
+
+    async fn save_index(&self, index: &ImageIndex) -> Result<()> {
+        let mut encoded = serde_json::to_string_pretty(index)?;
+        encoded.push('\n');
+        let mut index_path = self.cache_dir.clone();
+        index_path.push("index.json");
+        fs::write(&index_path, encoded).await?;
+        Ok(())
+    }
+}

crates/oci/src/packer/mod.rs

@@ -0,0 +1,69 @@
+use std::path::PathBuf;
+
+use crate::{name::ImageName, schema::OciSchema};
+
+use self::backend::OciPackerBackendType;
+use oci_spec::image::{Descriptor, ImageConfiguration, ImageManifest};
+
+pub mod backend;
+pub mod cache;
+pub mod service;
+
+#[derive(Debug, Default, Clone, Copy, Eq, PartialEq, Hash)]
+pub enum OciPackedFormat {
+    #[default]
+    Squashfs,
+    Erofs,
+    Tar,
+}
+
+impl OciPackedFormat {
+    pub fn extension(&self) -> &str {
+        match self {
+            OciPackedFormat::Squashfs => "squashfs",
+            OciPackedFormat::Erofs => "erofs",
+            OciPackedFormat::Tar => "tar",
+        }
+    }
+
+    pub fn backend(&self) -> OciPackerBackendType {
+        match self {
+            OciPackedFormat::Squashfs => OciPackerBackendType::MkSquashfs,
+            OciPackedFormat::Erofs => OciPackerBackendType::MkfsErofs,
+            OciPackedFormat::Tar => OciPackerBackendType::Tar,
+        }
+    }
+}
+
+#[derive(Clone)]
+pub struct OciPackedImage {
+    pub name: ImageName,
+    pub digest: String,
+    pub path: PathBuf,
+    pub format: OciPackedFormat,
+    pub descriptor: Descriptor,
+    pub config: OciSchema<ImageConfiguration>,
+    pub manifest: OciSchema<ImageManifest>,
+}
+
+impl OciPackedImage {
+    pub fn new(
+        name: ImageName,
+        digest: String,
+        path: PathBuf,
+        format: OciPackedFormat,
+        descriptor: Descriptor,
+        config: OciSchema<ImageConfiguration>,
+        manifest: OciSchema<ImageManifest>,
+    ) -> OciPackedImage {
+        OciPackedImage {
+            name,
+            digest,
+            path,
+            format,
+            descriptor,
+            config,
+            manifest,
+        }
+    }
+}

crates/oci/src/packer/service.rs

@@ -0,0 +1,278 @@
+use std::{
+    collections::{hash_map::Entry, HashMap},
+    fmt::Display,
+    path::{Path, PathBuf},
+    sync::Arc,
+};
+
+use anyhow::{anyhow, Result};
+use oci_spec::image::Descriptor;
+use tokio::{
+    sync::{watch, Mutex},
+    task::JoinHandle,
+};
+
+use crate::{
+    assemble::OciImageAssembler,
+    fetch::{OciImageFetcher, OciResolvedImage},
+    name::ImageName,
+    progress::{OciBoundProgress, OciProgress, OciProgressContext},
+    registry::OciPlatform,
+};
+
+use log::{error, info, warn};
+
+use super::{cache::OciPackerCache, OciPackedFormat, OciPackedImage};
+
+pub struct OciPackerTask {
+    progress: OciBoundProgress,
+    watch: watch::Sender<Option<Result<OciPackedImage>>>,
+    task: JoinHandle<()>,
+}
+
+#[derive(Clone)]
+pub struct OciPackerService {
+    seed: Option<PathBuf>,
+    platform: OciPlatform,
+    cache: OciPackerCache,
+    tasks: Arc<Mutex<HashMap<OciPackerTaskKey, OciPackerTask>>>,
+}
+
+impl OciPackerService {
+    pub async fn new(
+        seed: Option<PathBuf>,
+        cache_dir: &Path,
+        platform: OciPlatform,
+    ) -> Result<OciPackerService> {
+        Ok(OciPackerService {
+            seed,
+            cache: OciPackerCache::new(cache_dir).await?,
+            platform,
+            tasks: Arc::new(Mutex::new(HashMap::new())),
+        })
+    }
+
+    pub async fn list(&self) -> Result<Vec<Descriptor>> {
+        self.cache.list().await
+    }
+
+    pub async fn recall(
+        &self,
+        digest: &str,
+        format: OciPackedFormat,
+    ) -> Result<Option<OciPackedImage>> {
+        if digest.contains('/') || digest.contains('\\') || digest.contains("..") {
+            return Ok(None);
+        }
+
+        self.cache
+            .recall(ImageName::parse("cached:latest")?, digest, format)
+            .await
+    }
+
+    pub async fn request(
+        &self,
+        name: ImageName,
+        format: OciPackedFormat,
+        overwrite: bool,
+        progress_context: OciProgressContext,
+    ) -> Result<OciPackedImage> {
+        let progress = OciProgress::new();
+        let progress = OciBoundProgress::new(progress_context.clone(), progress);
+        let fetcher =
+            OciImageFetcher::new(self.seed.clone(), self.platform.clone(), progress.clone());
+        let resolved = fetcher.resolve(name.clone()).await?;
+        let key = OciPackerTaskKey {
+            digest: resolved.digest.clone(),
+            format,
+        };
+        let (progress_copy_task, mut receiver) = match self.tasks.lock().await.entry(key.clone()) {
+            Entry::Occupied(entry) => {
+                let entry = entry.get();
+                (
+                    Some(entry.progress.also_update(progress_context).await),
+                    entry.watch.subscribe(),
+                )
+            }
+
+            Entry::Vacant(entry) => {
+                let task = self
+                    .clone()
+                    .launch(
+                        name,
+                        key.clone(),
+                        format,
+                        overwrite,
+                        resolved,
+                        fetcher,
+                        progress.clone(),
+                    )
+                    .await;
+                let (watch, receiver) = watch::channel(None);
+
+                let task = OciPackerTask {
+                    progress: progress.clone(),
+                    task,
+                    watch,
+                };
+                entry.insert(task);
+                (None, receiver)
+            }
+        };
+
+        let _progress_task_guard = scopeguard::guard(progress_copy_task, |task| {
+            if let Some(task) = task {
+                task.abort();
+            }
+        });
+
+        let _task_cancel_guard = scopeguard::guard(self.clone(), |service| {
+            service.maybe_cancel_task(key);
+        });
+
+        loop {
+            receiver.changed().await?;
+            let current = receiver.borrow_and_update();
+            if current.is_some() {
+                return current
+                    .as_ref()
+                    .map(|x| x.as_ref().map_err(|err| anyhow!("{}", err)).cloned())
+                    .unwrap();
+            }
+        }
+    }
+
+    #[allow(clippy::too_many_arguments)]
+    async fn launch(
+        self,
+        name: ImageName,
+        key: OciPackerTaskKey,
+        format: OciPackedFormat,
+        overwrite: bool,
+        resolved: OciResolvedImage,
+        fetcher: OciImageFetcher,
+        progress: OciBoundProgress,
+    ) -> JoinHandle<()> {
+        info!("started packer task {}", key);
+        tokio::task::spawn(async move {
+            let _task_drop_guard =
+                scopeguard::guard((key.clone(), self.clone()), |(key, service)| {
+                    service.ensure_task_gone(key);
+                });
+            if let Err(error) = self
+                .task(
+                    name,
+                    key.clone(),
+                    format,
+                    overwrite,
+                    resolved,
+                    fetcher,
+                    progress,
+                )
+                .await
+            {
+                self.finish(&key, Err(error)).await;
+            }
+        })
+    }
+
+    #[allow(clippy::too_many_arguments)]
+    async fn task(
+        &self,
+        name: ImageName,
+        key: OciPackerTaskKey,
+        format: OciPackedFormat,
+        overwrite: bool,
+        resolved: OciResolvedImage,
+        fetcher: OciImageFetcher,
+        progress: OciBoundProgress,
+    ) -> Result<()> {
+        if !overwrite {
+            if let Some(cached) = self
+                .cache
+                .recall(name.clone(), &resolved.digest, format)
+                .await?
+            {
+                self.finish(&key, Ok(cached)).await;
+                return Ok(());
+            }
+        }
+        let assembler =
+            OciImageAssembler::new(fetcher, resolved, progress.clone(), None, None).await?;
+        let assembled = assembler.assemble().await?;
+        let mut file = assembled
+            .tmp_dir
+            .clone()
+            .ok_or(anyhow!("tmp_dir was missing when packing image"))?;
+        file.push("image.pack");
+        let target = file.clone();
+        let packer = format.backend().create();
+        packer
+            .pack(progress, assembled.vfs.clone(), &target)
+            .await?;
+        let packed = OciPackedImage::new(
+            name,
+            assembled.digest.clone(),
+            file,
+            format,
+            assembled.descriptor.clone(),
+            assembled.config.clone(),
+            assembled.manifest.clone(),
+        );
+        let packed = self.cache.store(packed).await?;
+        self.finish(&key, Ok(packed)).await;
+        Ok(())
+    }
+
+    async fn finish(&self, key: &OciPackerTaskKey, result: Result<OciPackedImage>) {
+        let Some(task) = self.tasks.lock().await.remove(key) else {
+            error!("packer task {} was not found when task completed", key);
+            return;
+        };
+
+        match result.as_ref() {
+            Ok(_) => {
+                info!("completed packer task {}", key);
+            }
+
+            Err(err) => {
+                warn!("packer task {} failed: {}", key, err);
+            }
+        }
+
+        task.watch.send_replace(Some(result));
+    }
+
+    fn maybe_cancel_task(self, key: OciPackerTaskKey) {
+        tokio::task::spawn(async move {
+            let tasks = self.tasks.lock().await;
+            if let Some(task) = tasks.get(&key) {
+                if task.watch.is_closed() {
+                    task.task.abort();
+                }
+            }
+        });
+    }
+
+    fn ensure_task_gone(self, key: OciPackerTaskKey) {
+        tokio::task::spawn(async move {
+            let mut tasks = self.tasks.lock().await;
+            if let Some(task) = tasks.remove(&key) {
+                warn!("aborted packer task {}", key);
+                task.watch.send_replace(Some(Err(anyhow!("task aborted"))));
+            }
+        });
+    }
+}
+
+#[derive(Debug, Clone, Eq, PartialEq, Hash)]
+struct OciPackerTaskKey {
+    digest: String,
+    format: OciPackedFormat,
+}
+
+impl Display for OciPackerTaskKey {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        f.write_fmt(format_args!("{}:{}", self.digest, self.format.extension()))
+    }
+}

crates/oci/src/progress.rs

@@ -0,0 +1,238 @@
+use indexmap::IndexMap;
+use std::sync::Arc;
+use tokio::{
+    sync::{watch, Mutex},
+    task::JoinHandle,
+};
+
+#[derive(Clone, Debug)]
+pub struct OciProgress {
+    pub phase: OciProgressPhase,
+    pub digest: Option<String>,
+    pub layers: IndexMap<String, OciProgressLayer>,
+    pub indication: OciProgressIndication,
+}
+
+impl Default for OciProgress {
+    fn default() -> Self {
+        Self::new()
+    }
+}
+
+impl OciProgress {
+    pub fn new() -> Self {
+        OciProgress {
+            phase: OciProgressPhase::Started,
+            digest: None,
+            layers: IndexMap::new(),
+            indication: OciProgressIndication::Hidden,
+        }
+    }
+
+    pub fn start_resolving(&mut self) {
+        self.phase = OciProgressPhase::Resolving;
+        self.indication = OciProgressIndication::Spinner { message: None };
+    }
+
+    pub fn resolved(&mut self, digest: &str) {
+        self.digest = Some(digest.to_string());
+        self.indication = OciProgressIndication::Hidden;
+    }
+
+    pub fn add_layer(&mut self, id: &str) {
+        self.layers.insert(
+            id.to_string(),
+            OciProgressLayer {
+                id: id.to_string(),
+                phase: OciProgressLayerPhase::Waiting,
+                indication: OciProgressIndication::Spinner { message: None },
+            },
+        );
+    }
+
+    pub fn downloading_layer(&mut self, id: &str, downloaded: u64, total: u64) {
+        if let Some(entry) = self.layers.get_mut(id) {
+            entry.phase = OciProgressLayerPhase::Downloading;
+            entry.indication = OciProgressIndication::ProgressBar {
+                message: None,
+                current: downloaded,
+                total,
+                bytes: true,
+            };
+        }
+    }
+
+    pub fn downloaded_layer(&mut self, id: &str, total: u64) {
+        if let Some(entry) = self.layers.get_mut(id) {
+            entry.phase = OciProgressLayerPhase::Downloaded;
+            entry.indication = OciProgressIndication::Completed {
+                message: None,
+                total: Some(total),
+                bytes: true,
+            };
+        }
+    }
+
+    pub fn start_assemble(&mut self) {
+        self.phase = OciProgressPhase::Assemble;
+        self.indication = OciProgressIndication::Hidden;
+    }
+
+    pub fn start_extracting_layer(&mut self, id: &str) {
+        if let Some(entry) = self.layers.get_mut(id) {
+            entry.phase = OciProgressLayerPhase::Extracting;
+            entry.indication = OciProgressIndication::Spinner { message: None };
+        }
+    }
+
+    pub fn extracting_layer(&mut self, id: &str, file: &str) {
+        if let Some(entry) = self.layers.get_mut(id) {
+            entry.phase = OciProgressLayerPhase::Extracting;
+            entry.indication = OciProgressIndication::Spinner {
+                message: Some(file.to_string()),
+            };
+        }
+    }
+
+    pub fn extracted_layer(&mut self, id: &str, count: u64, total_size: u64) {
+        if let Some(entry) = self.layers.get_mut(id) {
+            entry.phase = OciProgressLayerPhase::Extracted;
+            entry.indication = OciProgressIndication::Completed {
+                message: Some(format!("{} files", count)),
+                total: Some(total_size),
+                bytes: true,
+            };
+        }
+    }
+
+    pub fn start_packing(&mut self) {
+        self.phase = OciProgressPhase::Pack;
+        for layer in self.layers.values_mut() {
+            layer.indication = OciProgressIndication::Hidden;
+        }
+        self.indication = OciProgressIndication::Spinner { message: None };
+    }
+
+    pub fn complete(&mut self, size: u64) {
+        self.phase = OciProgressPhase::Complete;
+        self.indication = OciProgressIndication::Completed {
+            message: None,
+            total: Some(size),
+            bytes: true,
+        }
+    }
+}
+
+#[derive(Clone, Debug)]
+pub enum OciProgressPhase {
+    Started,
+    Resolving,
+    Resolved,
+    ConfigDownload,
+    LayerDownload,
+    Assemble,
+    Pack,
+    Complete,
+}
+
+#[derive(Clone, Debug)]
+pub enum OciProgressIndication {
+    Hidden,
+
+    ProgressBar {
+        message: Option<String>,
+        current: u64,
+        total: u64,
+        bytes: bool,
+    },
+
+    Spinner {
+        message: Option<String>,
+    },
+
+    Completed {
+        message: Option<String>,
+        total: Option<u64>,
+        bytes: bool,
+    },
+}
+
+#[derive(Clone, Debug)]
+pub struct OciProgressLayer {
+    pub id: String,
+    pub phase: OciProgressLayerPhase,
+    pub indication: OciProgressIndication,
+}
+
+#[derive(Clone, Debug)]
+pub enum OciProgressLayerPhase {
+    Waiting,
+    Downloading,
+    Downloaded,
+    Extracting,
+    Extracted,
+}
+
+#[derive(Clone)]
+pub struct OciProgressContext {
+    sender: watch::Sender<OciProgress>,
+}
+
+impl OciProgressContext {
+    pub fn create() -> (OciProgressContext, watch::Receiver<OciProgress>) {
+        let (sender, receiver) = watch::channel(OciProgress::new());
+        (OciProgressContext::new(sender), receiver)
+    }
+
+    pub fn new(sender: watch::Sender<OciProgress>) -> OciProgressContext {
+        OciProgressContext { sender }
+    }
+
+    pub fn update(&self, progress: &OciProgress) {
+        let _ = self.sender.send(progress.clone());
+    }
+
+    pub fn subscribe(&self) -> watch::Receiver<OciProgress> {
+        self.sender.subscribe()
+    }
+}
+
+#[derive(Clone)]
+pub struct OciBoundProgress {
+    context: OciProgressContext,
+    instance: Arc<Mutex<OciProgress>>,
+}
+
+impl OciBoundProgress {
+    pub fn new(context: OciProgressContext, progress: OciProgress) -> OciBoundProgress {
+        OciBoundProgress {
+            context,
+            instance: Arc::new(Mutex::new(progress)),
+        }
+    }
+
+    pub async fn update(&self, function: impl FnOnce(&mut OciProgress)) {
+        let mut progress = self.instance.lock().await;
+        function(&mut progress);
+        self.context.update(&progress);
+    }
+
+    pub fn update_blocking(&self, function: impl FnOnce(&mut OciProgress)) {
+        let mut progress = self.instance.blocking_lock();
+        function(&mut progress);
+        self.context.update(&progress);
+    }
+
+    pub async fn also_update(&self, context: OciProgressContext) -> JoinHandle<()> {
+        let progress = self.instance.lock().await.clone();
+        context.update(&progress);
+        let mut receiver = self.context.subscribe();
+        tokio::task::spawn(async move {
+            while (receiver.changed().await).is_ok() {
+                context
+                    .sender
+                    .send_replace(receiver.borrow_and_update().clone());
+            }
+        })
+    }
+}

crates/oci/src/registry.rs

@@ -7,26 +7,28 @@ use reqwest::{Client, RequestBuilder, Response, StatusCode};
 use tokio::{fs::File, io::AsyncWriteExt};
 use url::Url;
 
+use crate::{name::ImageName, progress::OciBoundProgress, schema::OciSchema};
+
 #[derive(Clone, Debug)]
-pub struct OciRegistryPlatform {
+pub struct OciPlatform {
     pub os: Os,
     pub arch: Arch,
 }
 
-impl OciRegistryPlatform {
+impl OciPlatform {
     #[cfg(target_arch = "x86_64")]
     const CURRENT_ARCH: Arch = Arch::Amd64;
     #[cfg(target_arch = "aarch64")]
     const CURRENT_ARCH: Arch = Arch::ARM64;
 
-    pub fn new(os: Os, arch: Arch) -> OciRegistryPlatform {
-        OciRegistryPlatform { os, arch }
+    pub fn new(os: Os, arch: Arch) -> OciPlatform {
+        OciPlatform { os, arch }
     }
 
-    pub fn current() -> OciRegistryPlatform {
-        OciRegistryPlatform {
+    pub fn current() -> OciPlatform {
+        OciPlatform {
             os: Os::Linux,
-            arch: OciRegistryPlatform::CURRENT_ARCH,
+            arch: OciPlatform::CURRENT_ARCH,
         }
     }
 }
@@ -34,12 +36,12 @@ impl OciRegistryPlatform {
 pub struct OciRegistryClient {
     agent: Client,
     url: Url,
-    platform: OciRegistryPlatform,
+    platform: OciPlatform,
     token: Option<String>,
 }
 
 impl OciRegistryClient {
-    pub fn new(url: Url, platform: OciRegistryPlatform) -> Result<OciRegistryClient> {
+    pub fn new(url: Url, platform: OciPlatform) -> Result<OciRegistryClient> {
         Ok(OciRegistryClient {
             agent: Client::new(),
             url,
@@ -138,6 +140,7 @@ impl OciRegistryClient {
         name: N,
         descriptor: &Descriptor,
         mut dest: File,
+        progress: Option<OciBoundProgress>,
     ) -> Result<u64> {
         let url = self.url.join(&format!(
             "/v2/{}/blobs/{}",
@@ -149,6 +152,18 @@ impl OciRegistryClient {
         while let Some(chunk) = response.chunk().await? {
             dest.write_all(&chunk).await?;
             size += chunk.len() as u64;
+
+            if let Some(ref progress) = progress {
+                progress
+                    .update(|progress| {
+                        progress.downloading_layer(
+                            descriptor.digest(),
+                            size,
+                            descriptor.size() as u64,
+                        );
+                    })
+                    .await;
+            }
         }
         Ok(size)
     }
@@ -157,11 +172,11 @@ impl OciRegistryClient {
         &mut self,
         name: N,
         reference: R,
-    ) -> Result<(ImageManifest, String)> {
+    ) -> Result<(OciSchema<ImageManifest>, String)> {
         let url = self.url.join(&format!(
             "/v2/{}/manifests/{}",
             name.as_ref(),
-            reference.as_ref()
+            reference.as_ref(),
         ))?;
         let accept = format!(
             "{}, {}, {}, {}",
@@ -179,20 +194,28 @@ impl OciRegistryClient {
             .ok_or_else(|| anyhow!("fetching manifest did not yield a content digest"))?
             .to_str()?
             .to_string();
-        let manifest = serde_json::from_str(&response.text().await?)?;
-        Ok((manifest, digest))
+        let bytes = response.bytes().await?;
+        let manifest = serde_json::from_slice(&bytes)?;
+        Ok((OciSchema::new(bytes.to_vec(), manifest), digest))
     }
 
     pub async fn get_manifest_with_digest<N: AsRef<str>, R: AsRef<str>>(
         &mut self,
         name: N,
-        reference: R,
-    ) -> Result<(ImageManifest, String)> {
-        let url = self.url.join(&format!(
-            "/v2/{}/manifests/{}",
-            name.as_ref(),
-            reference.as_ref()
-        ))?;
+        reference: Option<R>,
+        digest: Option<N>,
+    ) -> Result<(OciSchema<ImageManifest>, Option<Descriptor>, String)> {
+        let what = digest
+            .as_ref()
+            .map(|x| x.as_ref().to_string())
+            .unwrap_or_else(|| {
+                reference
+                    .map(|x| x.as_ref().to_string())
+                    .unwrap_or_else(|| ImageName::DEFAULT_IMAGE_TAG.to_string())
+            });
+        let url = self
+            .url
+            .join(&format!("/v2/{}/manifests/{}", name.as_ref(), what,))?;
         let accept = format!(
             "{}, {}, {}, {}",
             MediaType::ImageManifest.to_docker_v2s2()?,
@@ -215,18 +238,21 @@ impl OciRegistryClient {
             let descriptor = self
                 .pick_manifest(index)
                 .ok_or_else(|| anyhow!("unable to pick manifest from index"))?;
-            return self
+            let (manifest, digest) = self
                 .get_raw_manifest_with_digest(name, descriptor.digest())
-                .await;
+                .await?;
+            return Ok((manifest, Some(descriptor), digest));
         }
         let digest = response
             .headers()
             .get("Docker-Content-Digest")
-            .ok_or_else(|| anyhow!("fetching manifest did not yield a content digest"))?
-            .to_str()?
-            .to_string();
-        let manifest = serde_json::from_str(&response.text().await?)?;
-        Ok((manifest, digest))
+            .and_then(|x| x.to_str().ok())
+            .map(|x| x.to_string())
+            .or_else(|| digest.map(|x: N| x.as_ref().to_string()))
+            .ok_or_else(|| anyhow!("fetching manifest did not yield a content digest"))?;
+        let bytes = response.bytes().await?;
+        let manifest = serde_json::from_slice(&bytes)?;
+        Ok((OciSchema::new(bytes.to_vec(), manifest), None, digest))
     }
 
     fn pick_manifest(&mut self, index: ImageIndex) -> Option<Descriptor> {

crates/oci/src/schema.rs

@@ -0,0 +1,29 @@
+use std::fmt::Debug;
+
+#[derive(Clone, Debug)]
+pub struct OciSchema<T: Clone + Debug> {
+    raw: Vec<u8>,
+    item: T,
+}
+
+impl<T: Clone + Debug> OciSchema<T> {
+    pub fn new(raw: Vec<u8>, item: T) -> OciSchema<T> {
+        OciSchema { raw, item }
+    }
+
+    pub fn raw(&self) -> &[u8] {
+        &self.raw
+    }
+
+    pub fn item(&self) -> &T {
+        &self.item
+    }
+
+    pub fn into_raw(self) -> Vec<u8> {
+        self.raw
+    }
+
+    pub fn into_item(self) -> T {
+        self.item
+    }
+}

crates/oci/src/vfs.rs

@@ -0,0 +1,264 @@
+use std::path::{Path, PathBuf};
+
+use anyhow::{anyhow, Result};
+use tokio::{
+    fs::File,
+    io::{AsyncRead, AsyncWrite, AsyncWriteExt},
+};
+use tokio_tar::{Builder, Entry, EntryType, Header};
+
+#[derive(Clone, Debug, Eq, PartialEq)]
+pub enum VfsNodeType {
+    Directory,
+    RegularFile,
+    Symlink,
+    Hardlink,
+    Fifo,
+    CharDevice,
+    BlockDevice,
+}
+
+#[derive(Clone, Debug)]
+pub struct VfsNode {
+    pub name: String,
+    pub size: u64,
+    pub children: Vec<VfsNode>,
+    pub typ: VfsNodeType,
+    pub uid: u64,
+    pub gid: u64,
+    pub link_name: Option<String>,
+    pub mode: u32,
+    pub mtime: u64,
+    pub dev_major: Option<u32>,
+    pub dev_minor: Option<u32>,
+    pub disk_path: Option<PathBuf>,
+}
+
+impl VfsNode {
+    pub fn from<X: AsyncRead + Unpin>(entry: &Entry<X>) -> Result<VfsNode> {
+        let header = entry.header();
+        let name = entry
+            .path()?
+            .file_name()
+            .ok_or(anyhow!("unable to get file name for entry"))?
+            .to_string_lossy()
+            .to_string();
+        let typ = header.entry_type();
+        let vtype = if typ.is_symlink() {
+            VfsNodeType::Symlink
+        } else if typ.is_hard_link() {
+            VfsNodeType::Hardlink
+        } else if typ.is_dir() {
+            VfsNodeType::Directory
+        } else if typ.is_fifo() {
+            VfsNodeType::Fifo
+        } else if typ.is_block_special() {
+            VfsNodeType::BlockDevice
+        } else if typ.is_character_special() {
+            VfsNodeType::CharDevice
+        } else if typ.is_file() {
+            VfsNodeType::RegularFile
+        } else {
+            return Err(anyhow!("unable to determine vfs type for entry"));
+        };
+
+        Ok(VfsNode {
+            name,
+            size: header.size()?,
+            children: vec![],
+            typ: vtype,
+            uid: header.uid()?,
+            gid: header.gid()?,
+            link_name: header.link_name()?.map(|x| x.to_string_lossy().to_string()),
+            mode: header.mode()?,
+            mtime: header.mtime()?,
+            dev_major: header.device_major()?,
+            dev_minor: header.device_minor()?,
+            disk_path: None,
+        })
+    }
+
+    pub fn lookup(&self, path: &Path) -> Option<&VfsNode> {
+        let mut node = self;
+        for part in path {
+            node = node
+                .children
+                .iter()
+                .find(|child| child.name == part.to_string_lossy())?;
+        }
+        Some(node)
+    }
+
+    pub fn lookup_mut(&mut self, path: &Path) -> Option<&mut VfsNode> {
+        let mut node = self;
+        for part in path {
+            node = node
+                .children
+                .iter_mut()
+                .find(|child| child.name == part.to_string_lossy())?;
+        }
+        Some(node)
+    }
+
+    pub fn remove(&mut self, path: &Path) -> Option<(&mut VfsNode, VfsNode)> {
+        let parent = path.parent()?;
+        let node = self.lookup_mut(parent)?;
+        let file_name = path.file_name()?;
+        let file_name = file_name.to_string_lossy();
+        let position = node
+            .children
+            .iter()
+            .position(|child| file_name == child.name)?;
+        let removed = node.children.remove(position);
+        Some((node, removed))
+    }
+
+    pub fn create_tar_header(&self) -> Result<Header> {
+        let mut header = Header::new_ustar();
+        header.set_entry_type(match self.typ {
+            VfsNodeType::Directory => EntryType::Directory,
+            VfsNodeType::CharDevice => EntryType::Char,
+            VfsNodeType::BlockDevice => EntryType::Block,
+            VfsNodeType::Fifo => EntryType::Fifo,
+            VfsNodeType::Hardlink => EntryType::Link,
+            VfsNodeType::Symlink => EntryType::Symlink,
+            VfsNodeType::RegularFile => EntryType::Regular,
+        });
+        header.set_uid(self.uid);
+        header.set_gid(self.gid);
+
+        if let Some(device_major) = self.dev_major {
+            header.set_device_major(device_major)?;
+        }
+
+        if let Some(device_minor) = self.dev_minor {
+            header.set_device_minor(device_minor)?;
+        }
+        header.set_mtime(self.mtime);
+        header.set_mode(self.mode);
+
+        if let Some(link_name) = self.link_name.as_ref() {
+            header.set_link_name(&PathBuf::from(link_name))?;
+        }
+        header.set_size(self.size);
+        Ok(header)
+    }
+
+    pub async fn write_to_tar<W: AsyncWrite + Unpin + Send>(
+        &self,
+        path: &Path,
+        builder: &mut Builder<W>,
+    ) -> Result<()> {
+        let mut header = self.create_tar_header()?;
+        header.set_path(path)?;
+        header.set_cksum();
+        if let Some(disk_path) = self.disk_path.as_ref() {
+            builder
+                .append(&header, File::open(disk_path).await?)
+                .await?;
+        } else {
+            builder.append(&header, &[] as &[u8]).await?;
+        }
+        Ok(())
+    }
+}
+
+#[derive(Clone, Debug)]
+pub struct VfsTree {
+    pub root: VfsNode,
+}
+
+impl Default for VfsTree {
+    fn default() -> Self {
+        Self::new()
+    }
+}
+
+impl VfsTree {
+    pub fn new() -> VfsTree {
+        VfsTree {
+            root: VfsNode {
+                name: "".to_string(),
+                size: 0,
+                children: vec![],
+                typ: VfsNodeType::Directory,
+                uid: 0,
+                gid: 0,
+                link_name: None,
+                mode: 0,
+                mtime: 0,
+                dev_major: None,
+                dev_minor: None,
+                disk_path: None,
+            },
+        }
+    }
+
+    pub fn insert_tar_entry<X: AsyncRead + Unpin>(&mut self, entry: &Entry<X>) -> Result<&VfsNode> {
+        let mut meta = VfsNode::from(entry)?;
+        let path = entry.path()?.to_path_buf();
+        let parent = if let Some(parent) = path.parent() {
+            self.root.lookup_mut(parent)
+        } else {
+            Some(&mut self.root)
+        };
+
+        let Some(parent) = parent else {
+            return Err(anyhow!("unable to find parent of entry"));
+        };
+
+        let position = parent
+            .children
+            .iter()
+            .position(|child| meta.name == child.name);
+
+        if let Some(position) = position {
+            let old = parent.children.remove(position);
+            if meta.typ == VfsNodeType::Directory {
+                meta.children = old.children;
+            }
+        }
+        parent.children.push(meta.clone());
+        let Some(reference) = parent.children.iter().find(|child| child.name == meta.name) else {
+            return Err(anyhow!("unable to find inserted child in vfs"));
+        };
+        Ok(reference)
+    }
+
+    pub fn set_disk_path(&mut self, path: &Path, disk_path: &Path) -> Result<()> {
+        let Some(node) = self.root.lookup_mut(path) else {
+            return Err(anyhow!(
+                "unable to find node {:?} to set disk path to",
+                path
+            ));
+        };
+        node.disk_path = Some(disk_path.to_path_buf());
+        Ok(())
+    }
+
+    pub async fn write_to_tar<W: AsyncWrite + Unpin + Send + 'static>(
+        &self,
+        write: W,
+    ) -> Result<()> {
+        let mut builder = Builder::new(write);
+        let mut queue = vec![(PathBuf::from(""), &self.root)];
+
+        while !queue.is_empty() {
+            let (mut path, node) = queue.remove(0);
+            if !node.name.is_empty() {
+                path.push(&node.name);
+            }
+            if path.components().count() != 0 {
+                node.write_to_tar(&path, &mut builder).await?;
+            }
+            for child in &node.children {
+                queue.push((path.clone(), child));
+            }
+        }
+
+        let mut write = builder.into_inner().await?;
+        write.flush().await?;
+        drop(write);
+        Ok(())
+    }
+}

crates/runtime/Cargo.toml

@@ -2,7 +2,7 @@
 name = "krata-runtime"
 description = "Runtime for running guests on the krata hypervisor."
 license.workspace = true
-version= "0.0.3"
+version.workspace = true
 homepage.workspace = true
 repository.workspace = true
 edition = "2021"
@@ -12,18 +12,18 @@ resolver = "2"
 anyhow = { workspace = true }
 backhand = { workspace = true }
 ipnetwork = { workspace = true }
-krata = { path = "../krata", version = "^0.0.3" }
+krata = { path = "../krata", version = "^0.0.10" }
 krata-advmac = { workspace = true }
-krata-oci = { path = "../oci", version = "^0.0.3" }
+krata-oci = { path = "../oci", version = "^0.0.10" }
 log = { workspace = true }
 loopdev-3 = { workspace = true }
 serde_json = { workspace = true }
 tokio = { workspace = true }
 uuid = { workspace = true }
-krata-xenclient = { path = "../xen/xenclient", version = "^0.0.3" }
-krata-xenevtchn = { path = "../xen/xenevtchn", version = "^0.0.3" }
-krata-xengnt = { path = "../xen/xengnt", version = "^0.0.3" }
-krata-xenstore = { path = "../xen/xenstore", version = "^0.0.3" }
+krata-xenclient = { path = "../xen/xenclient", version = "^0.0.10" }
+krata-xenevtchn = { path = "../xen/xenevtchn", version = "^0.0.10" }
+krata-xengnt = { path = "../xen/xengnt", version = "^0.0.10" }
+krata-xenstore = { path = "../xen/xenstore", version = "^0.0.10" }
 
 [lib]
 name = "kratart"
@@ -31,10 +31,6 @@ name = "kratart"
 [dev-dependencies]
 env_logger = { workspace = true }
 
-[[example]]
-name = "kratart-squashify"
-path = "examples/squashify.rs"
-
 [[example]]
 name = "kratart-channel"
 path = "examples/channel.rs"

crates/runtime/examples/channel.rs

@@ -6,7 +6,7 @@ use kratart::channel::ChannelService;
 async fn main() -> Result<()> {
     env_logger::Builder::from_env(Env::default().default_filter_or("info")).init();
 
-    let (service, mut receiver) = ChannelService::new("krata-channel".to_string()).await?;
+    let (service, _, mut receiver) = ChannelService::new("krata-channel".to_string(), None).await?;
     let task = service.launch().await?;
 
     loop {

crates/runtime/examples/squashify.rs

@@ -1,29 +0,0 @@
-use std::{env::args, path::PathBuf};
-
-use anyhow::Result;
-use env_logger::Env;
-use krataoci::{cache::ImageCache, compiler::ImageCompiler, name::ImageName};
-use tokio::fs;
-
-#[tokio::main]
-async fn main() -> Result<()> {
-    env_logger::Builder::from_env(Env::default().default_filter_or("info")).init();
-
-    let image = ImageName::parse(&args().nth(1).unwrap())?;
-    let seed = args().nth(2).map(PathBuf::from);
-
-    let cache_dir = PathBuf::from("krata-cache");
-    if !cache_dir.exists() {
-        fs::create_dir(&cache_dir).await?;
-    }
-
-    let cache = ImageCache::new(&cache_dir)?;
-    let compiler = ImageCompiler::new(&cache, seed)?;
-    let info = compiler.compile(&image).await?;
-    println!(
-        "generated squashfs of {} to {}",
-        image,
-        info.image_squashfs.to_string_lossy()
-    );
-    Ok(())
-}

crates/runtime/src/autoloop.rs

@@ -1,17 +1,25 @@
+use std::{sync::Arc, time::Duration};
+
 use anyhow::{anyhow, Result};
+use log::debug;
 use loopdev::{LoopControl, LoopDevice};
+use tokio::time::sleep;
 use xenclient::BlockDeviceRef;
 
+#[derive(Clone)]
 pub struct AutoLoop {
-    control: LoopControl,
+    control: Arc<LoopControl>,
 }
 
 impl AutoLoop {
     pub fn new(control: LoopControl) -> AutoLoop {
-        AutoLoop { control }
+        AutoLoop {
+            control: Arc::new(control),
+        }
     }
 
     pub fn loopify(&self, file: &str) -> Result<BlockDeviceRef> {
+        debug!("creating loop for file {}", file);
         let device = self.control.next_free()?;
         device.with().read_only(true).attach(file)?;
         let path = device
@@ -25,9 +33,10 @@ impl AutoLoop {
         Ok(BlockDeviceRef { path, major, minor })
     }
 
-    pub fn unloop(&self, device: &str) -> Result<()> {
+    pub async fn unloop(&self, device: &str) -> Result<()> {
         let device = LoopDevice::open(device)?;
         device.detach()?;
+        sleep(Duration::from_millis(200)).await;
         Ok(())
     }
 }

crates/runtime/src/cfgblk.rs

@@ -1,7 +1,7 @@
 use anyhow::Result;
 use backhand::{FilesystemWriter, NodeHeader};
 use krata::launchcfg::LaunchInfo;
-use krataoci::compiler::ImageInfo;
+use krataoci::packer::OciPackedImage;
 use log::trace;
 use std::fs;
 use std::fs::File;
@@ -9,28 +9,24 @@ use std::path::PathBuf;
 use uuid::Uuid;
 
 pub struct ConfigBlock<'a> {
-    pub image_info: &'a ImageInfo,
+    pub image: &'a OciPackedImage,
     pub file: PathBuf,
     pub dir: PathBuf,
 }
 
 impl ConfigBlock<'_> {
-    pub fn new<'a>(uuid: &Uuid, image_info: &'a ImageInfo) -> Result<ConfigBlock<'a>> {
+    pub fn new<'a>(uuid: &Uuid, image: &'a OciPackedImage) -> Result<ConfigBlock<'a>> {
         let mut dir = std::env::temp_dir().clone();
         dir.push(format!("krata-cfg-{}", uuid));
         fs::create_dir_all(&dir)?;
         let mut file = dir.clone();
         file.push("config.squashfs");
-        Ok(ConfigBlock {
-            image_info,
-            file,
-            dir,
-        })
+        Ok(ConfigBlock { image, file, dir })
     }
 
     pub fn build(&self, launch_config: &LaunchInfo) -> Result<()> {
         trace!("build launch_config={:?}", launch_config);
-        let manifest = self.image_info.config.to_string()?;
+        let config = self.image.config.raw();
         let launch = serde_json::to_string(launch_config)?;
         let mut writer = FilesystemWriter::default();
         writer.push_dir(
@@ -43,7 +39,7 @@ impl ConfigBlock<'_> {
             },
         )?;
         writer.push_file(
-            manifest.as_bytes(),
+            config,
             "/image/config.json",
             NodeHeader {
                 permissions: 384,

crates/runtime/src/channel.rs

@@ -41,30 +41,40 @@ impl XenConsoleInterface {
 
 pub struct ChannelService {
     typ: String,
+    use_reserved_ref: Option<u64>,
     backends: HashMap<u32, ChannelBackend>,
     evtchn: EventChannel,
     store: XsdClient,
     gnttab: GrantTab,
     input_receiver: Receiver<(u32, Vec<u8>)>,
     pub input_sender: Sender<(u32, Vec<u8>)>,
-    output_sender: Sender<(u32, Vec<u8>)>,
+    output_sender: Sender<(u32, Option<Vec<u8>>)>,
 }
 
 impl ChannelService {
-    pub async fn new(typ: String) -> Result<(ChannelService, Receiver<(u32, Vec<u8>)>)> {
+    pub async fn new(
+        typ: String,
+        use_reserved_ref: Option<u64>,
+    ) -> Result<(
+        ChannelService,
+        Sender<(u32, Vec<u8>)>,
+        Receiver<(u32, Option<Vec<u8>>)>,
+    )> {
         let (input_sender, input_receiver) = channel(GROUPED_CHANNEL_QUEUE_LEN);
         let (output_sender, output_receiver) = channel(GROUPED_CHANNEL_QUEUE_LEN);
         Ok((
             ChannelService {
                 typ,
+                use_reserved_ref,
                 backends: HashMap::new(),
                 evtchn: EventChannel::open().await?,
                 store: XsdClient::open().await?,
                 gnttab: GrantTab::open()?,
-                input_sender,
+                input_sender: input_sender.clone(),
                 input_receiver,
                 output_sender,
             },
+            input_sender,
             output_receiver,
         ))
     }
@@ -148,6 +158,7 @@ impl ChannelService {
             self.evtchn.clone(),
             self.gnttab.clone(),
             self.output_sender.clone(),
+            self.use_reserved_ref,
         )
         .await?;
         self.backends.insert(domid, backend);
@@ -192,12 +203,14 @@ pub struct ChannelBackend {
     pub domid: u32,
     pub id: u32,
     pub sender: Sender<Vec<u8>>,
+    raw_sender: Sender<(u32, Option<Vec<u8>>)>,
     task: JoinHandle<()>,
 }
 
 impl Drop for ChannelBackend {
     fn drop(&mut self) {
         self.task.abort();
+        let _ = self.raw_sender.try_send((self.domid, None));
         debug!(
             "destroyed channel backend for domain {} channel {}",
             self.domid, self.id
@@ -215,7 +228,8 @@ impl ChannelBackend {
         store: XsdClient,
         evtchn: EventChannel,
         gnttab: GrantTab,
-        output_sender: Sender<(u32, Vec<u8>)>,
+        output_sender: Sender<(u32, Option<Vec<u8>>)>,
+        use_reserved_ref: Option<u64>,
     ) -> Result<ChannelBackend> {
         let processor = KrataChannelBackendProcessor {
             backend,
@@ -225,15 +239,19 @@ impl ChannelBackend {
             store,
             evtchn,
             gnttab,
+            use_reserved_ref,
         };
 
         let (input_sender, input_receiver) = channel(SINGLE_CHANNEL_QUEUE_LEN);
 
-        let task = processor.launch(output_sender, input_receiver).await?;
+        let task = processor
+            .launch(output_sender.clone(), input_receiver)
+            .await?;
         Ok(ChannelBackend {
             domid,
             id,
             task,
+            raw_sender: output_sender,
             sender: input_sender,
         })
     }
@@ -241,6 +259,7 @@ impl ChannelBackend {
 
 #[derive(Clone)]
 pub struct KrataChannelBackendProcessor {
+    use_reserved_ref: Option<u64>,
     backend: String,
     frontend: String,
     id: u32,
@@ -290,7 +309,7 @@ impl KrataChannelBackendProcessor {
 
     async fn launch(
         &self,
-        output_sender: Sender<(u32, Vec<u8>)>,
+        output_sender: Sender<(u32, Option<Vec<u8>>)>,
         input_receiver: Receiver<Vec<u8>>,
     ) -> Result<JoinHandle<()>> {
         let owned = self.clone();
@@ -307,7 +326,7 @@ impl KrataChannelBackendProcessor {
 
     async fn processor(
         &self,
-        sender: Sender<(u32, Vec<u8>)>,
+        sender: Sender<(u32, Option<Vec<u8>>)>,
         mut receiver: Receiver<Vec<u8>>,
     ) -> Result<()> {
         self.init().await?;
@@ -332,7 +351,7 @@ impl KrataChannelBackendProcessor {
                                 .read_string(format!("{}/port", self.frontend))
                                 .await?;
 
-                            if (ring_ref.is_none() || port.is_none()) && tries < 10 {
+                            if (ring_ref.is_none() || port.is_none()) && tries < 40 {
                                 tries += 1;
                                 self.store
                                     .write_string(format!("{}/state", self.backend), "4")
@@ -347,7 +366,7 @@ impl KrataChannelBackendProcessor {
                             return Err(anyhow!("frontend did not give ring-ref and port"));
                         }
 
-                        let Ok(ring_ref) = ring_ref.unwrap().parse::<u64>() else {
+                        let Ok(mut ring_ref) = ring_ref.unwrap().parse::<u64>() else {
                             return Err(anyhow!("frontend gave invalid ring-ref"));
                         };
 
@@ -355,6 +374,8 @@ impl KrataChannelBackendProcessor {
                             return Err(anyhow!("frontend gave invalid port"));
                         };
 
+                        ring_ref = self.use_reserved_ref.unwrap_or(ring_ref);
+
                         break (ring_ref, port);
                     }
                 }
@@ -380,7 +401,7 @@ impl KrataChannelBackendProcessor {
         unsafe {
             let buffer = self.read_output_buffer(channel.local_port, &memory).await?;
             if !buffer.is_empty() {
-                sender.send((self.domid, buffer)).await?;
+                sender.send((self.domid, Some(buffer))).await?;
             }
         };
 
@@ -427,6 +448,10 @@ impl KrataChannelBackendProcessor {
                                 error!("channel for domid {} has an invalid input space of {}", self.domid, space);
                             }
                             let free = XenConsoleInterface::INPUT_SIZE.wrapping_sub(space);
+                            if free == 0 {
+                                sleep(Duration::from_micros(100)).await;
+                                continue;
+                            }
                             let want = data.len().min(free);
                             let buffer = &data[index..want];
                             for b in buffer {
@@ -450,7 +475,7 @@ impl KrataChannelBackendProcessor {
                         unsafe {
                             let buffer = self.read_output_buffer(channel.local_port, &memory).await?;
                             if !buffer.is_empty() {
-                                sender.send((self.domid, buffer)).await?;
+                                sender.send((self.domid, Some(buffer))).await?;
                             }
                         };
                         channel.unmask_sender.send(channel.local_port).await?;

crates/runtime/src/console.rs

@@ -1,18 +0,0 @@
-use anyhow::Result;
-use tokio::fs::File;
-
-pub struct XenConsole {
-    pub read_handle: File,
-    pub write_handle: File,
-}
-
-impl XenConsole {
-    pub async fn new(tty: &str) -> Result<XenConsole> {
-        let read_handle = File::options().read(true).write(false).open(tty).await?;
-        let write_handle = File::options().read(false).write(true).open(tty).await?;
-        Ok(XenConsole {
-            read_handle,
-            write_handle,
-        })
-    }
-}

crates/runtime/src/launch.rs

@@ -1,5 +1,6 @@
 use std::collections::HashMap;
 use std::net::{IpAddr, Ipv6Addr};
+use std::sync::Arc;
 use std::{fs, net::Ipv4Addr, str::FromStr};
 
 use advmac::MacAddr6;
@@ -7,48 +8,49 @@ use anyhow::{anyhow, Result};
 use ipnetwork::{IpNetwork, Ipv4Network};
 use krata::launchcfg::{
     LaunchInfo, LaunchNetwork, LaunchNetworkIpv4, LaunchNetworkIpv6, LaunchNetworkResolver,
+    LaunchPackedFormat, LaunchRoot,
 };
+use krataoci::packer::OciPackedImage;
+use tokio::sync::Semaphore;
 use uuid::Uuid;
 use xenclient::{DomainChannel, DomainConfig, DomainDisk, DomainNetworkInterface};
 use xenstore::XsdInterface;
 
 use crate::cfgblk::ConfigBlock;
 use crate::RuntimeContext;
-use krataoci::{
-    cache::ImageCache,
-    compiler::{ImageCompiler, ImageInfo},
-    name::ImageName,
-};
 
 use super::{GuestInfo, GuestState};
 
-pub struct GuestLaunchRequest<'a> {
+pub struct GuestLaunchRequest {
+    pub format: LaunchPackedFormat,
+    pub kernel: Vec<u8>,
+    pub initrd: Vec<u8>,
     pub uuid: Option<Uuid>,
-    pub name: Option<&'a str>,
-    pub image: &'a str,
+    pub name: Option<String>,
     pub vcpus: u32,
     pub mem: u64,
     pub env: HashMap<String, String>,
     pub run: Option<Vec<String>>,
     pub debug: bool,
+    pub image: OciPackedImage,
 }
 
-pub struct GuestLauncher {}
+pub struct GuestLauncher {
+    pub launch_semaphore: Arc<Semaphore>,
+}
 
 impl GuestLauncher {
-    pub fn new() -> Result<Self> {
-        Ok(Self {})
+    pub fn new(launch_semaphore: Arc<Semaphore>) -> Result<Self> {
+        Ok(Self { launch_semaphore })
     }
 
-    pub async fn launch<'r>(
+    pub async fn launch(
         &mut self,
-        context: &mut RuntimeContext,
-        request: GuestLaunchRequest<'r>,
+        context: &RuntimeContext,
+        request: GuestLaunchRequest,
     ) -> Result<GuestInfo> {
         let uuid = request.uuid.unwrap_or_else(Uuid::new_v4);
         let xen_name = format!("krata-{uuid}");
-        let image_info = self.compile(request.image, &context.image_cache).await?;
-
         let mut gateway_mac = MacAddr6::random();
         gateway_mac.set_local(true);
         gateway_mac.set_multicast(false);
@@ -56,6 +58,7 @@ impl GuestLauncher {
         container_mac.set_local(true);
         container_mac.set_multicast(false);
 
+        let _launch_permit = self.launch_semaphore.acquire().await?;
         let guest_ipv4 = self.allocate_ipv4(context).await?;
         let guest_ipv6 = container_mac.to_link_local_ipv6();
         let gateway_ipv4 = "10.75.70.1";
@@ -64,6 +67,16 @@ impl GuestLauncher {
         let ipv6_network_mask: u32 = 10;
 
         let launch_config = LaunchInfo {
+            root: LaunchRoot {
+                format: request.format.clone(),
+            },
+            hostname: Some(
+                request
+                    .name
+                    .as_ref()
+                    .map(|x| x.to_string())
+                    .unwrap_or_else(|| format!("krata-{}", uuid)),
+            ),
             network: Some(LaunchNetwork {
                 link: "eth0".to_string(),
                 ipv4: LaunchNetworkIpv4 {
@@ -87,13 +100,14 @@ impl GuestLauncher {
             run: request.run,
         };
 
-        let cfgblk = ConfigBlock::new(&uuid, &image_info)?;
+        let cfgblk = ConfigBlock::new(&uuid, &request.image)?;
         cfgblk.build(&launch_config)?;
 
-        let image_squashfs_path = image_info
-            .image_squashfs
+        let image_squashfs_path = request
+            .image
+            .path
             .to_str()
-            .ok_or_else(|| anyhow!("failed to convert image squashfs path to string"))?;
+            .ok_or_else(|| anyhow!("failed to convert image path to string"))?;
 
         let cfgblk_dir_path = cfgblk
             .dir
@@ -129,7 +143,6 @@ impl GuestLauncher {
                     cfgblk_dir_path,
                 ),
             ),
-            ("krata/image".to_string(), request.image.to_string()),
             (
                 "krata/network/guest/ipv4".to_string(),
                 format!("{}/{}", guest_ipv4, ipv4_network_mask),
@@ -156,27 +169,28 @@ impl GuestLauncher {
             ),
         ];
 
-        if let Some(name) = request.name {
-            extra_keys.push(("krata/name".to_string(), name.to_string()));
+        if let Some(name) = request.name.as_ref() {
+            extra_keys.push(("krata/name".to_string(), name.clone()));
         }
 
         let config = DomainConfig {
             backend_domid: 0,
-            name: &xen_name,
+            name: xen_name,
             max_vcpus: request.vcpus,
             mem_mb: request.mem,
-            kernel_path: &context.kernel,
-            initrd_path: &context.initrd,
-            cmdline: &cmdline,
+            kernel: request.kernel,
+            initrd: request.initrd,
+            cmdline,
+            use_console_backend: Some("krata-console".to_string()),
             disks: vec![
                 DomainDisk {
-                    vdev: "xvda",
-                    block: &image_squashfs_loop,
+                    vdev: "xvda".to_string(),
+                    block: image_squashfs_loop.clone(),
                     writable: false,
                 },
                 DomainDisk {
-                    vdev: "xvdb",
-                    block: &cfgblk_squashfs_loop,
+                    vdev: "xvdb".to_string(),
+                    block: cfgblk_squashfs_loop.clone(),
                     writable: false,
                 },
             ],
@@ -185,7 +199,7 @@ impl GuestLauncher {
                 initialized: false,
             }],
             vifs: vec![DomainNetworkInterface {
-                mac: &guest_mac_string,
+                mac: guest_mac_string.clone(),
                 mtu: 1500,
                 bridge: None,
                 script: None,
@@ -197,10 +211,10 @@ impl GuestLauncher {
         };
         match context.xen.create(&config).await {
             Ok(created) => Ok(GuestInfo {
-                name: request.name.map(|x| x.to_string()),
+                name: request.name.as_ref().map(|x| x.to_string()),
                 uuid,
                 domid: created.domid,
-                image: request.image.to_string(),
+                image: request.image.digest,
                 loops: vec![],
                 guest_ipv4: Some(IpNetwork::new(
                     IpAddr::V4(guest_ipv4),
@@ -217,27 +231,21 @@ impl GuestLauncher {
                 )?),
                 gateway_ipv6: Some(IpNetwork::new(
                     IpAddr::V6(Ipv6Addr::from_str(gateway_ipv6)?),
-                    ipv4_network_mask as u8,
+                    ipv6_network_mask as u8,
                 )?),
                 gateway_mac: Some(gateway_mac_string.clone()),
                 state: GuestState { exit_code: None },
             }),
             Err(error) => {
-                let _ = context.autoloop.unloop(&image_squashfs_loop.path);
-                let _ = context.autoloop.unloop(&cfgblk_squashfs_loop.path);
+                let _ = context.autoloop.unloop(&image_squashfs_loop.path).await;
+                let _ = context.autoloop.unloop(&cfgblk_squashfs_loop.path).await;
                 let _ = fs::remove_dir(&cfgblk.dir);
                 Err(error.into())
             }
         }
     }
 
-    async fn compile(&self, image: &str, image_cache: &ImageCache) -> Result<ImageInfo> {
-        let image = ImageName::parse(image)?;
-        let compiler = ImageCompiler::new(image_cache, None)?;
-        compiler.compile(&image).await
-    }
-
-    async fn allocate_ipv4(&mut self, context: &mut RuntimeContext) -> Result<Ipv4Addr> {
+    async fn allocate_ipv4(&self, context: &RuntimeContext) -> Result<Ipv4Addr> {
         let network = Ipv4Network::new(Ipv4Addr::new(10, 75, 80, 0), 24)?;
         let mut used: Vec<Ipv4Addr> = vec![];
         for domid_candidate in context.xen.store.list("/local/domain").await? {
@@ -264,7 +272,7 @@ impl GuestLauncher {
 
         if found.is_none() {
             return Err(anyhow!(
-                "unable to find ipv4 to allocate to container, ipv4 addresses are exhausted"
+                "unable to find ipv4 to allocate to guest, ipv4 addresses are exhausted"
             ));
         }