264 lines
21 KiB
Markdown
264 lines
21 KiB
Markdown
# Vendored `containerization` — Nucleic patches
|
||
|
||
> Overview of *why* these patches exist (the session-isolation model) + the build/validate workflow:
|
||
> [`docs/CONTAINER_ISOLATION.md`](../../docs/CONTAINER_ISOLATION.md). This file is the per-patch detail.
|
||
|
||
This is a **vendored copy** of [apple/containerization](https://github.com/apple/containerization)
|
||
at upstream commit `6b7b42ca3efeee8c706070e4355e6a807c5336ae`, referenced by the root `Package.swift`
|
||
via `.package(path: "third_party/containerization")` instead of the github URL.
|
||
|
||
It is vendored (not pulled) because we carry a local patch upstream doesn't have. Keeping it
|
||
in-tree means the patch can't be lost to a dependency re-resolve.
|
||
|
||
## What's changed vs. upstream
|
||
|
||
1. **`Sources/Containerization/LinuxContainer.swift` — forward VM extensions.**
|
||
`LinuxContainer.Configuration` gains a `vmExtensions: [any Sendable]` field, and
|
||
`LinuxContainer` assigns it into `VMConfiguration.extensions` when it builds the VM config.
|
||
Upstream already supports `VMConfiguration.extensions` + the `VZInstanceExtension` hook
|
||
(`configureVZ`/`didCreate`), but `LinuxContainer` — the only entry point we use — never forwarded
|
||
it, so there was no way to attach a device (e.g. a virtio memory balloon) to a container's VM.
|
||
Search for the marker comment `[Nucleic vendored patch]` to find both edit sites.
|
||
|
||
Nucleic uses this to attach a `VZVirtioTraditionalMemoryBalloonDeviceConfiguration` and drive its
|
||
target at runtime for automatic VM memory reclamation — see `MemoryBalloon.swift` /
|
||
`ContainerEngine` in NucleicCore.
|
||
|
||
2. **`Sources/Containerization/LinuxProcess.swift` — process-group kill.**
|
||
`LinuxProcess` gains `killProcessGroup(_:)`, which signals the negative pid (`-pid`) so the
|
||
guest's `kill(2)` targets the exec'd process's whole **process group**, not just the leader.
|
||
Every exec is `setsid()`'d by `vmexec`, so the process is its own group leader (pgid == pid) and
|
||
a group signal reaches the children it forked. Upstream only exposes the leader-only `kill(_:)`,
|
||
which let a forked child survive a Stop in a long-lived shared container. Marked with
|
||
`[Nucleic vendored patch]`; used by `ContainerizedProcessHandle.sendSignal` in NucleicCore.
|
||
|
||
3. **`Sources/Containerization/LinuxProcess.swift` — stdio-connection diagnostics (log-only).**
|
||
`setupIO` logs (`os.Logger`, subsystem `com.nucleic`, category `container-io`) when a *configured*
|
||
stdio stream's guest side never connects — which leaves its host `FileHandle` nil, so the relay /
|
||
readability handler is never wired and the agent's stdin is never delivered (it hangs) or its
|
||
stdout is never read (the "no output, just a spinner" symptom in Nucleic Control containers).
|
||
Behavior is unchanged; it only surfaces the failing stream. Marked `[Nucleic vendored patch]`
|
||
(the `import os`, the `nucleicIOLog` static, and the per-stream check in `setupIO`). All three are
|
||
wrapped in `#if canImport(os)` — non-Xcode toolchains (e.g. a swiftly Swift used to cross-build the
|
||
host framework) resolve Foundation/Virtualization but not the `os` overlay, so the diagnostic
|
||
degrades to a no-op there instead of failing the build; Xcode (the local `make vminit-image` path)
|
||
builds keep it.
|
||
|
||
4. **Trimmed for footprint (no behavior change).** `Tests/`, `docs/`, `examples/`, and `images/`
|
||
were dropped, and the corresponding `.testTarget(...)` entries removed from `Package.swift`. The
|
||
library/executable targets we build are untouched.
|
||
|
||
5. **`Sources/Containerization/LinuxProcess.swift` — non-blocking stdio relay.**
|
||
Upstream's `setupIO` relays guest stdout/stderr with `FileHandle.availableData`, a **blocking**
|
||
read, from inside a `readabilityHandler`. Those handlers run on Foundation's shared readability
|
||
queue, so if one exec's guest stdout wedged mid-stream that blocking read parked the shared thread
|
||
and head-of-line-blocked **every** other exec's stdout/stderr relay across all containers — one
|
||
stuck session froze the others. The patch marks each connected fd `O_NONBLOCK` and drains it via a
|
||
new `nucleicDrainNonBlocking` (returns bytes + EOF, never blocks; EAGAIN just waits for the next
|
||
readable event). A wedged stream is now contained to its own exec. Marked `[Nucleic vendored patch]`
|
||
(the two static helpers `nucleicSetNonBlocking`/`nucleicDrainNonBlocking` and the two rewritten
|
||
`readabilityHandler` blocks). Requires host-side POSIX `read`/`fcntl`/`errno`.
|
||
|
||
6. **`Sources/Containerization/LinuxProcess.swift` — atomic stdio-or-abort start.**
|
||
In `start()`, after `setupIO` returns, if a *configured* stdio stream never connected from the
|
||
guest (its `FileHandle` is nil — patch #3's logged failure), the patch tears the just-created exec
|
||
back down (`agent.deleteProcess`) and throws instead of calling `startProcess`. Upstream proceeds
|
||
and runs a process with a dead stream (stdin never delivered → hangs; stdout never read → the "no
|
||
output, just a spinner" 60s stall in Nucleic Control). Now that permanent silent stall surfaces as
|
||
a clean, retryable start error. Marked `[Nucleic vendored patch]` (the guard block before
|
||
`startProcess`).
|
||
|
||
7. **`Sources/Containerization/Vminitd.swift` — bounded teardown RPC.**
|
||
`deleteProcess` now sends a 30s `CallOptions.timeout` (upstream sends none, so it can block
|
||
forever on a wedged agent channel). Nucleic calls `LinuxProcess.delete()` after every turn to
|
||
reclaim the per-exec vsock/gRPC connection `exec()` dials; an unbounded `deleteProcess` would let
|
||
that reclaim hang and the connection leak. On the thrown deadline, `performDeletion` still closes
|
||
the agent connection. Marked `[Nucleic vendored patch]` (the `callOpts` block in `deleteProcess`).
|
||
NOTE: this pairs with a Nucleic-side change in `ContainerizedProcessHandle` (call `delete()` after
|
||
the exec exits / on force-close) — without that caller, upstream never deletes execs at all and
|
||
the shared control container leaks a connection + `runConnections()` task per turn.
|
||
|
||
10. **`Sources/ContainerizationOS/Socket/Socket.swift` — `acceptStream` survives transient accept
|
||
errors.** Upstream cancelled the accept `DispatchSource` on ANY `accept(2)` failure, permanently
|
||
ending accepting while the socket stayed **bound and listening** — a silent black hole: every
|
||
later client `connect(2)` SUCCEEDED into the kernel backlog and hung forever unanswered. When
|
||
the socket is the relayed control plane of a shared container, that is the "agent produced no
|
||
output within 60s / stdio transport stalled" all-sessions wedge, fixable only by recreating the
|
||
VM (an app restart). Transient errors — `ECONNABORTED`/`ECONNRESET` (a queued connection dying
|
||
before accept, routine under connection churn), `EMFILE`/`ENFILE`/`ENOBUFS`/`ENOMEM` (resource
|
||
pressure), `EINTR`/`EAGAIN` — now skip that one accept and keep listening (new
|
||
`isTransientAcceptError`). This is SHARED code: the fix reaches the host by a normal build and
|
||
the guest via the initfs rebuild. Marked `[Nucleic vendored patch]`.
|
||
|
||
11. **`Sources/Containerization/UnixSocketRelay.swift` — relay loops contain per-connection
|
||
failures.** Both accept loops (`setupHostVsockListener` — the host half of a container's relayed
|
||
control socket — and `setupHostVsockDial`) used to let ONE thrown per-connection dial/connect
|
||
(host server rebinding, backlog momentarily full → `ECONNREFUSED`, fd pressure) propagate out of
|
||
the loop, whose teardown then removed the vsock listener — permanently severing every session in
|
||
the container from the host control plane. Each connection is now handled on its own task with
|
||
its error logged and contained (mirroring the guest `VsockProxy`), and the pre-relay failure
|
||
paths close both ends so a failed connection fails FAST for the peer and leaks no fds. Marked
|
||
`[Nucleic vendored patch]`.
|
||
|
||
13. **`Sources/ContainerizationOS/Keychain/KeychainQuery.swift` — prompt-free registry-credential
|
||
reads.** Upstream's `get`/`list`/`exists` call `SecItemCopyMatching` with the legacy login
|
||
Keychain's interactive authorization panel enabled, so any process that isn't on a registry
|
||
internet-password item's ACL raises the macOS *"'cctl' wants to use your confidential information
|
||
stored in 'ghcr.io' in your keychain"* panel when it reads that item — e.g. a `cctl` binary
|
||
re-signed ad-hoc by a fresh `make vminit-image` reading a GHCR token an earlier build stored, or
|
||
any tool linking `KeychainHelper.lookup` during an image pull/push/list. Nucleic's rule is that no
|
||
automatic credential lookup may ever raise a Keychain panel. This patch wraps the three
|
||
`SecItemCopyMatching` reads in `withoutInteractiveUI` (`SecKeychainSetUserInteractionAllowed(false)`
|
||
— the only switch that governs the legacy ACL/partition-list dialog; the data-protection
|
||
`kSecUseAuthenticationUI*` flags do NOT), so an already-trusted item reads silently while anything
|
||
else fails with `errSecInteractionNotAllowed` — which `isQuerySuccessful` now treats as "not found"
|
||
so the caller falls back to anonymous / `REGISTRY_HOST`/`USERNAME`/`TOKEN` env auth. `save`
|
||
(the `cctl login` write path) gains a delete-and-retry on `errSecDuplicateItem`, since the
|
||
now-silent `exists` can under-report an unreadable pre-existing item. Mirrors
|
||
`KeychainOwnedAccess.withoutLegacyKeychainUI` in NucleicCore. Host-side (shipped by a normal
|
||
`swift build`). The three `SecKeychain*` symbols are formally deprecated but are the only API
|
||
covering the legacy ACL panel; they're bound directly via `@_silgen_name` (`nucleic_SecKeychain*`,
|
||
top of file) so the required calls compile without deprecation warnings. Marked
|
||
`[Nucleic vendored patch]`.
|
||
|
||
14. **`Sources/Containerization/Vminitd.swift` — configure the gRPC pipeline before the channel goes
|
||
active.** `Vminitd.init` used the now-deprecated `HTTP2ClientTransport.WrappedChannel.wrapping(
|
||
channel:config:serviceConfig:)`, which wraps an already-connected channel best-effort and may drop
|
||
early server frames such as SETTINGS. Migrated to `wrapping(config:serviceConfig:makeChannel:)`,
|
||
which invokes the transport's `configure` inside the bootstrap's channel initializer — before the
|
||
vsock channel becomes active. `init` is now `async throws` (the new overload is async); its callers
|
||
in `VZVirtualMachineInstance` (`start`/`dialAgent`, both already async) and the integration test now
|
||
`try await`. Marked `[Nucleic vendored patch]`.
|
||
|
||
### GUEST-side patches (require rebuilding the initfs — see below)
|
||
|
||
Patches #1–#7 are host-side (the `Containerization` library), shipped by a normal `swift build`.
|
||
Patches #8+ live in `vminitd/` (the guest agent), which rides in the initfs OCI image. They are INERT
|
||
until that image is rebuilt from this source and published, and `ContainerEngine.vminitReference`
|
||
points at it. Build it with **`make vminit-image`** (root Makefile) — it builds cctl + the guest
|
||
vminitd/vmexec from this vendored tree and packages `ghcr.io/abkslm/vminit:<tag>` into the local cctl
|
||
store; `make vminit-image-push` publishes it (authenticate once with `make vminit-image-login`, which
|
||
stores a GHCR token in the macOS Keychain — or set `REGISTRY_HOST`/`USERNAME`/`TOKEN`), and
|
||
`vminitReference` is pinned to that custom image. First time on a machine, run `make vminit-image-prep` once (installs
|
||
the swiftly toolchain + musl SDK the guest cross-build needs). Bump the `-nucleicN` tag suffix and
|
||
rebuild whenever a guest patch changes. Built locally, not in CI: the host framework needs the macOS
|
||
26+ Virtualization SDK that GitHub-hosted runners lack.
|
||
|
||
8. **`vminitd/Sources/VminitdCore/ManagedProcess.swift` — offload the blocking start off the event loop.**
|
||
`ManagedProcess.start()` did synchronous, potentially slow pipe reads (waiting for `vmexec` to
|
||
return the pid, then for the error pipe to close) while holding `state`'s Mutex, ON the calling
|
||
task — which is the gRPC handler's event-loop thread. A slow start therefore parked the loop and
|
||
head-of-line-blocked sibling execs' control RPCs sharing it. The patch splits the body into a
|
||
synchronous `startBlocking()` and an async `start()` that runs it on `DispatchQueue.global` via a
|
||
checked continuation, keeping the loop responsive. Safe because the body has no `await` and
|
||
`ManagedProcess` is `Sendable`. Marked `[Nucleic vendored patch]`.
|
||
|
||
9. **Per-exec cgroups (OOM/CPU/pids isolation).** Upstream puts the container init AND every exec in
|
||
ONE cgroup (`/container/<id>`), so one session's runaway RSS trips the in-VM OOM-killer against a
|
||
*random* sibling, and a fork bomb / CPU hog hits the whole box. This patch makes `/container/<id>`
|
||
an intermediary: the resource ceiling stays on it, `ManagedContainer.init` moves the init into its
|
||
own leaf (`/container/<id>/init`) which enables `cgroup.subtree_control` up the chain, and
|
||
`ManagedProcess.start` places each exec in its OWN child (`/container/<id>/<execID>`) with
|
||
`memory.oom.group=1` (a runaway session's OOM kills only *its* tree), a fair `cpu.weight`, and a
|
||
`pids.max` fork-bomb backstop. New `Cgroup2Manager` helpers: `setOomGroup`/`setCpuWeight`/
|
||
`setPidsMax`/`remove`. **Best-effort with a graceful fallback**: if any step of the per-exec setup
|
||
fails it wipes the partial state and reverts to the flat layout, and `ManagedProcess` falls back to
|
||
the container cgroup per exec — so a cgroup hiccup degrades to today's behavior, never a failed
|
||
start. `ManagedContainer.execCgroupParent == nil` marks flat mode.
|
||
|
||
Beyond scoped-OOM, a **hard host-configured per-exec `memory.max`** is also wired — WITHOUT a
|
||
protobuf change, because the exec already ships the full OCI `Spec` and the guest just ignored
|
||
`linux.resources`. Host: `LinuxProcessConfiguration.memoryLimitInBytes` → `LinuxContainer.exec`
|
||
stamps it onto `spec.linux.resources.memory.limit`. Guest: `Server+GRPC.createProcess` reads that
|
||
back and passes it to `createExec`/`ManagedProcess`, which sets `memory.max` (new
|
||
`Cgroup2Manager.setMemoryMax`) on the exec's cgroup — so a session can't consume the whole box
|
||
before its own OOM. Driven by Nucleic's `ContainerServiceSettings.controlPerSessionMemoryGiB`
|
||
(default 0 = off; applied only to the shared control container, via `ContainerManager.exec`), so the
|
||
default stays scoped-OOM-only. Marked `[Nucleic vendored patch]` across `Cgroup2Manager.swift`,
|
||
`ManagedContainer.swift`, `ManagedProcess.swift`, `Server+GRPC.swift` (guest) and
|
||
`LinuxProcessConfiguration.swift`, `LinuxContainer.swift` (host). **COMPILE-VERIFIED (host + musl
|
||
cross-build); NOT yet runtime-validated** — a
|
||
wrong cgroup-v2 hierarchy fails at runtime, so boot a container with the new image and confirm
|
||
sessions start, `/sys/fs/cgroup/container/<id>/<execID>` exists per session, and a hog is contained,
|
||
before pointing a shipping build at it. `vmexec/RunCommand` is unchanged: it still applies
|
||
`linux.resources` at `linux.cgroupsPath`, which the patch repoints (init leaf) and clears
|
||
accordingly.
|
||
|
||
12. **`vminitd/Sources/VminitdCore/VsockProxy.swift` — leak-proof, crash-proof relay connections;
|
||
no black-hole listener.** Four fixes to the guest half of the relayed control socket (the path
|
||
every session's MCP/approval traffic crosses in a shared container):
|
||
- **fd leak (the root of the recurring all-sessions stall):** `cleanup` ran its two epoll
|
||
unregisters and two `close(2)`s in one `do/catch`, so a thrown unregister SKIPPED the closes —
|
||
leaking both connection fds. Control-plane traffic is connection-churny by design (an SSE
|
||
`tools/call` closes its connection every gated call; every intercepted git/gh/command event is
|
||
a short-lived connection), so the leaks accumulated until vminitd hit `EMFILE`, its accept
|
||
path began failing, and — before patch #10 — the accept stream died with the guest socket
|
||
still bound: every session in the container then stalled ("produced no output within 60s")
|
||
until the VM was recreated. Each cleanup step now runs independently.
|
||
- **double-resume crash:** both fds' epoll handlers can reach the cleanup condition; a second
|
||
entry would resume the `CheckedContinuation` twice — a fatal trap in the VM's PID-1 agent.
|
||
`cleanup` is now once-guarded.
|
||
- **`try!` registrations:** an `epoll_ctl` failure crashed vminitd outright; registration
|
||
failures now fail only that connection, releasing whatever was already set up.
|
||
- **no black-hole listener:** if the accept loop ever ends unexpectedly, the proxy now closes
|
||
its listener (new `listenerLoopEnded`), so peers get fail-fast refusals instead of connecting
|
||
into a never-accepted backlog. A failed pre-relay connection is also closed explicitly.
|
||
Marked `[Nucleic vendored patch]`.
|
||
|
||
14. **Non-blocking, non-spinning guest I/O plane (`IOPair.swift`, `OSFile+Splice.swift`,
|
||
`VsockProxy.swift`) — the root cause of the "control plane unresponsive / all sessions stall"
|
||
wedge.** Every exec's stdio relay (`IOPair`) and every control-plane relay connection
|
||
(`VsockProxy`) share ONE thread: `ProcessSupervisor.default`'s epoll poller. Three defects let
|
||
a single slow peer freeze that thread — and with it every session's stdio AND the whole
|
||
container's control plane at once (probes then read "accepts connections but never answers";
|
||
before the manager-side recovery rework the host restarted the container over this, SIGKILLing
|
||
every session):
|
||
- **`IOPair` blocking write:** only *registered* fds get `O_NONBLOCK` (set by `Epoll.add`), and
|
||
the relay registers only its READ fd — the write fd (e.g. the vsock socket carrying an exec's
|
||
stdout to the host) stayed blocking. One slow-drained stream parked the poller thread in
|
||
`write(2)`. The relay now sets the write fd non-blocking up front and implements real
|
||
backpressure: a full destination stashes the remainder in a `pending` backlog, registers the
|
||
write fd for EPOLLOUT, and suspends reads until the flush completes (throttling the producing
|
||
process via its own pipe, not the shared thread). The old close-on-short-write path — dead
|
||
while the fd was blocking, live and stdio-dropping once non-blocking — is subsumed by the
|
||
backlog; genuine write errors still close the pair.
|
||
- **`OSFile.splice` busy-spin:** when the destination was full (EAGAIN) and the source idle,
|
||
the outer loop had no exit — it spun the poller thread at 100% until the peer drained (or
|
||
forever if it never did). The flush leg's EAGAIN branch now returns partial progress; the
|
||
un-flushed bytes stay in the transfer pipe and the destination's EPOLLOUT edge resumes the
|
||
flush (both `VsockProxy` handlers already pump both directions on both events).
|
||
- **`VsockProxy` registration race:** the client fd's epoll handler can fire — and splice
|
||
toward the server fd — before the second registration makes that fd non-blocking; a full
|
||
destination made that a genuinely blocking splice on the poller thread. Both fds are now set
|
||
non-blocking before either is registered.
|
||
All three are guest-side and INERT until the initfs image is rebuilt (`make vminit-image`, tag
|
||
`0.34.0-nucleic4`) and `ContainerEngine.vminitReference` is bumped after runtime validation.
|
||
Marked `[Nucleic vendored patch]`.
|
||
|
||
## Re-vendoring a newer upstream commit
|
||
|
||
1. `git clone` upstream (or copy `.build/checkouts/containerization` after bumping the URL pin
|
||
temporarily), check out the desired commit.
|
||
2. `rsync -a --exclude=.git --exclude=.build --exclude=.swiftpm --exclude=Tests/ --exclude=docs/ \
|
||
--exclude=images/ <upstream>/ third_party/containerization/`
|
||
3. Remove the `.testTarget(...)` blocks from `third_party/containerization/Package.swift`.
|
||
4. Re-apply patch #1 (the `vmExtensions` field + the `vmConfig.extensions = …` forward), patch #2
|
||
(`LinuxProcess.killProcessGroup(_:)`), patch #3 (the `setupIO` stdio-connection log + its
|
||
`import os` / `nucleicIOLog`), patch #5 (the non-blocking stdio relay: `nucleicSetNonBlocking` /
|
||
`nucleicDrainNonBlocking` + the rewritten `readabilityHandler` blocks), and patch #6 (the atomic
|
||
stdio-or-abort guard in `start()`), patch #7 (the bounded `deleteProcess` timeout in
|
||
`Vminitd.swift`), and patch #8 (the `ManagedProcess.start` event-loop offload in `vminitd/`). Grep
|
||
for `[Nucleic vendored patch]` to find every site, and patch #9 (per-exec cgroups) across
|
||
`Cgroup2Manager.swift` / `ManagedContainer.swift` / `ManagedProcess.swift`, patch #10
|
||
(`Socket.acceptStream` transient-error tolerance + `isTransientAcceptError`), patch #11 (the
|
||
`UnixSocketRelay` per-connection containment + fail-fast closes), patch #12 (the `VsockProxy`
|
||
cleanup/`try!`/listener hardening in `vminitd/`), patch #13 (the prompt-free
|
||
`KeychainQuery` reads: `withoutInteractiveUI` + the `errSecInteractionNotAllowed` handling +
|
||
the `save` duplicate retry), and patch #14 (the non-blocking/non-spinning guest I/O plane:
|
||
`IOPair` backpressure, the `OSFile.splice` EAGAIN return, and the `VsockProxy` pre-registration
|
||
non-blocking fds — all in `vminitd/`). After re-applying any
|
||
`vminitd/` patch, rebuild + publish the custom init image
|
||
with `make vminit-image` + `make vminit-image-push`, and bump `ContainerEngine.vminitReference`.
|
||
5. Update the commit hash above and in the root `Package.swift` comment.
|
||
6. `swift build` and run the balloon tests.
|