Fix rootfs race condition causing missing podman#4
Merged
Conversation
When multiple VM tests run in parallel, a race condition occurs: 1. First process creates rootfs file via dd 2. Second process checks if file exists, sees it, uses it 3. But first process hasn't finished installing packages yet Result: VMs start with rootfs that has no podman installed. Fix: Create rootfs at temp path (base.ext4.tmp), then rename to final path (base.ext4) only after package installation completes. This ensures the file check is atomic with completion.
Rootfs creation in CI can take 2+ minutes due to: - Cloud image download (~7 seconds) - virt-customize (~10-60+ seconds, variable) - NBD extraction (~30 seconds) - Package installation (~60 seconds) Increase timeouts: - test_sanity: 180 -> 300 seconds - test_egress: 60 -> 180 seconds (baseline) - test_exec: 60 -> 180 seconds Clone VMs keep 60s timeout since they use cached rootfs.
Include the actual error message from Firecracker when API calls fail. This helps diagnose issues like 400 Bad Request during snapshot load.
When poll_health_by_pid times out, kill the baseline VM process before returning the error. This releases any locks held by that process (e.g., rootfs creation lock). Also increase timeout from 180 to 300 seconds to match fresh test.
- Add rust-toolchain.toml to lock Rust version (prevents format drift) - Add dependabot.yml for weekly dependency updates - Apply rustfmt 1.92.0 formatting fixes
Required for network_overrides support in snapshot load API. This enables clone VMs to use different TAP devices from baseline.
- Add find_firecracker() that validates version before returning path - Fail early with clear error if Firecracker is too old - Required for network_overrides support in snapshot cloning
The BuildJet runners have systemd-resolved holding port 53, which prevents dnsmasq from starting. This caused DNS resolution failures in the VM Egress bridged clone tests. Stop systemd-resolved before installing/starting dnsmasq.
Health check fixes (rootless mode): - Use nsenter to curl guest directly instead of slirp4netns port forwarding - slirp4netns port forwarding cannot reach addresses outside its 10.0.2.0/24 network, so DNAT approach failed - Health checks now enter the namespace via nsenter and curl 192.168.1.2:80 - Simplified slirp.rs: removed DNAT rules, kept MASQUERADE for egress DNS fixes: - Add get_host_dns_servers() to read DNS from /etc/resolv.conf - Falls back to /run/systemd/resolve/resolv.conf for systemd-resolved - Remove dnsmasq setup from CI workflows (no longer needed) - Update bridged.rs to use host DNS instead of dnsmasq Other: - Lower minimum Firecracker version to 1.13.1 for CI compatibility
VMs now use host DNS servers directly (read from /etc/resolv.conf) instead of relying on dnsmasq. This simplifies the network setup and removes the need to wait for dnsmasq to bind to the veth IP.
The VM tests (Sanity, Exec, Egress) all share /dev/nbd0 via bind mount and flock doesn't work across podman containers. When they ran in parallel, multiple virt-customize processes tried to access the same cloud image simultaneously, causing one to hang indefinitely. Fix by adding job dependencies so they run sequentially: - VM Sanity runs first - VM Exec waits for Sanity - VM Egress waits for Exec Added `if: always()` so later jobs still run even if earlier ones fail, since the rootfs will be cached after the first successful creation.
ejc3
added a commit
that referenced
this pull request
Feb 23, 2026
…hang Replace the racy double-rebind approach with a deterministic handshake chain that guarantees the exec server's AsyncFd epoll is re-registered before the host starts health-checking. Reduces restore-to-healthy from ~61s to ~0.5s. ## The Problem After snapshot restore, Firecracker's vsock transport reset (VIRTIO_VSOCK_EVENT_TRANSPORT_RESET) leaves the exec server's AsyncFd epoll registration stale. The previous fix (c15aa6b) removed the duplicate rebind signal from agent.rs but left a timing gap: if the restore-epoch watcher's single signal arrived late, the host's health monitor would start exec calls against a stale listener, hanging for ~60s until the kernel's vsock cleanup expired the stale connections. ## Trace Evidence (the smoking gun) From the vsock muxer log of a failing run (vm-ba97c): T+0.009s Exec call #1 → WORKS (167+144+176+123+71+27 bytes response) T+0.076s Exec call #2 → WORKS T+0.520s Exec call #3 → guest ACKs, receives request, sends NOTHING → 5s timeout T+5.5-55s Exec calls #4-#9 → same pattern: kernel accepts, app never processes T+60.5s Guest sends RST for ALL stale connections simultaneously T+60.5s Exec call #10 → WORKS → "container running status running=true" The container started at T+0.28s. The exec server was broken for 60 more seconds because the duplicate re_register() from agent.rs corrupted the edge-triggered epoll: the old AsyncFd consumed the edge notification, and the new AsyncFd never received events for pending connections. ## The Fix: Deterministic Handshake Chain exec_rebind_signal → exec_re_register → rebind_done → output.reconnect() ↓ host accepts output connection ↓ health monitor spawns Every transition has an explicit signal. Zero timing dependencies. ### fc-agent side (4 files): - exec.rs: After re_register(), signals rebind_done (AtomicBool + Notify) - restore.rs: Signals exec rebind, waits for rebind_done confirmation (5s timeout), THEN reconnects output vsock - agent.rs: Removed duplicate rebind signal after notify_cache_ready_and_wait; added exec_rebind_done/exec_rebind_done_notify Arcs - mmds.rs: Threads new params through watch_restore_epoch to both handle_clone_restore call sites ### Host side (3 files): - listeners.rs: Added connected_tx oneshot to run_output_listener(), fired on first output connection accept - snapshot.rs: Waits for output_connected_rx (30s timeout) before spawning health monitor; removed stale output_reconnect.notify_one() for startup snapshots - podman/mod.rs: Passes None for connected_tx (non-snapshot path) ## Results Before: restore-to-healthy = ~61s (exec broken, 9 consecutive 5s timeouts) After: restore-to-healthy = ~0.5s (35ms to output connected, 533ms to healthy) Post-restore exec stress test: 10 parallel calls completed in 16.3ms (max single: 15.3ms), zero timeouts. Tested: make test-root FILTER=localhost_rootless_btrfs_snapshot_restore STREAM=1
ejc3
added a commit
that referenced
this pull request
Mar 2, 2026
Fix rootfs race condition causing missing podman
ejc3
added a commit
that referenced
this pull request
Mar 2, 2026
…hang Replace the racy double-rebind approach with a deterministic handshake chain that guarantees the exec server's AsyncFd epoll is re-registered before the host starts health-checking. Reduces restore-to-healthy from ~61s to ~0.5s. ## The Problem After snapshot restore, Firecracker's vsock transport reset (VIRTIO_VSOCK_EVENT_TRANSPORT_RESET) leaves the exec server's AsyncFd epoll registration stale. The previous fix (c15aa6b) removed the duplicate rebind signal from agent.rs but left a timing gap: if the restore-epoch watcher's single signal arrived late, the host's health monitor would start exec calls against a stale listener, hanging for ~60s until the kernel's vsock cleanup expired the stale connections. ## Trace Evidence (the smoking gun) From the vsock muxer log of a failing run (vm-ba97c): T+0.009s Exec call #1 → WORKS (167+144+176+123+71+27 bytes response) T+0.076s Exec call #2 → WORKS T+0.520s Exec call #3 → guest ACKs, receives request, sends NOTHING → 5s timeout T+5.5-55s Exec calls #4-#9 → same pattern: kernel accepts, app never processes T+60.5s Guest sends RST for ALL stale connections simultaneously T+60.5s Exec call #10 → WORKS → "container running status running=true" The container started at T+0.28s. The exec server was broken for 60 more seconds because the duplicate re_register() from agent.rs corrupted the edge-triggered epoll: the old AsyncFd consumed the edge notification, and the new AsyncFd never received events for pending connections. ## The Fix: Deterministic Handshake Chain exec_rebind_signal → exec_re_register → rebind_done → output.reconnect() ↓ host accepts output connection ↓ health monitor spawns Every transition has an explicit signal. Zero timing dependencies. ### fc-agent side (4 files): - exec.rs: After re_register(), signals rebind_done (AtomicBool + Notify) - restore.rs: Signals exec rebind, waits for rebind_done confirmation (5s timeout), THEN reconnects output vsock - agent.rs: Removed duplicate rebind signal after notify_cache_ready_and_wait; added exec_rebind_done/exec_rebind_done_notify Arcs - mmds.rs: Threads new params through watch_restore_epoch to both handle_clone_restore call sites ### Host side (3 files): - listeners.rs: Added connected_tx oneshot to run_output_listener(), fired on first output connection accept - snapshot.rs: Waits for output_connected_rx (30s timeout) before spawning health monitor; removed stale output_reconnect.notify_one() for startup snapshots - podman/mod.rs: Passes None for connected_tx (non-snapshot path) ## Results Before: restore-to-healthy = ~61s (exec broken, 9 consecutive 5s timeouts) After: restore-to-healthy = ~0.5s (35ms to output connected, 533ms to healthy) Post-restore exec stress test: 10 parallel calls completed in 16.3ms (max single: 15.3ms), zero timeouts. Tested: make test-root FILTER=localhost_rootless_btrfs_snapshot_restore STREAM=1
ejc3
added a commit
that referenced
this pull request
Mar 2, 2026
Fix rootfs race condition causing missing podman
ejc3
added a commit
that referenced
this pull request
Mar 2, 2026
…hang Replace the racy double-rebind approach with a deterministic handshake chain that guarantees the exec server's AsyncFd epoll is re-registered before the host starts health-checking. Reduces restore-to-healthy from ~61s to ~0.5s. ## The Problem After snapshot restore, Firecracker's vsock transport reset (VIRTIO_VSOCK_EVENT_TRANSPORT_RESET) leaves the exec server's AsyncFd epoll registration stale. The previous fix (c15aa6b) removed the duplicate rebind signal from agent.rs but left a timing gap: if the restore-epoch watcher's single signal arrived late, the host's health monitor would start exec calls against a stale listener, hanging for ~60s until the kernel's vsock cleanup expired the stale connections. ## Trace Evidence (the smoking gun) From the vsock muxer log of a failing run (vm-ba97c): T+0.009s Exec call #1 → WORKS (167+144+176+123+71+27 bytes response) T+0.076s Exec call #2 → WORKS T+0.520s Exec call #3 → guest ACKs, receives request, sends NOTHING → 5s timeout T+5.5-55s Exec calls #4-#9 → same pattern: kernel accepts, app never processes T+60.5s Guest sends RST for ALL stale connections simultaneously T+60.5s Exec call #10 → WORKS → "container running status running=true" The container started at T+0.28s. The exec server was broken for 60 more seconds because the duplicate re_register() from agent.rs corrupted the edge-triggered epoll: the old AsyncFd consumed the edge notification, and the new AsyncFd never received events for pending connections. ## The Fix: Deterministic Handshake Chain exec_rebind_signal → exec_re_register → rebind_done → output.reconnect() ↓ host accepts output connection ↓ health monitor spawns Every transition has an explicit signal. Zero timing dependencies. ### fc-agent side (4 files): - exec.rs: After re_register(), signals rebind_done (AtomicBool + Notify) - restore.rs: Signals exec rebind, waits for rebind_done confirmation (5s timeout), THEN reconnects output vsock - agent.rs: Removed duplicate rebind signal after notify_cache_ready_and_wait; added exec_rebind_done/exec_rebind_done_notify Arcs - mmds.rs: Threads new params through watch_restore_epoch to both handle_clone_restore call sites ### Host side (3 files): - listeners.rs: Added connected_tx oneshot to run_output_listener(), fired on first output connection accept - snapshot.rs: Waits for output_connected_rx (30s timeout) before spawning health monitor; removed stale output_reconnect.notify_one() for startup snapshots - podman/mod.rs: Passes None for connected_tx (non-snapshot path) ## Results Before: restore-to-healthy = ~61s (exec broken, 9 consecutive 5s timeouts) After: restore-to-healthy = ~0.5s (35ms to output connected, 533ms to healthy) Post-restore exec stress test: 10 parallel calls completed in 16.3ms (max single: 15.3ms), zero timeouts. Tested: make test-root FILTER=localhost_rootless_btrfs_snapshot_restore STREAM=1
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
When multiple VM tests run in parallel (VM Exec, VM Egress, VM Sanity), a race condition occurs:
ddduringextract_root_partition()install_packages_in_rootfs()yetResult: VMs start with a rootfs that has no podman installed, causing:
Fix
Create rootfs at temp path (
base.ext4.tmp), then rename to final path (base.ext4) only after all package installation completes. This makes the existence check atomic with completion.Also fixes
The fuser fork has a formatting fix pushed (import ordering in integration_tests.rs).