egress-proxies-operator/docs/plans-executions/2026-09-17-1515-kvm2-review-fixes.md

Execution: fix the kvm2 code-review findings

Plan: docs/plans/2026-09-17-1515-kvm2-review-fixes.md

• Step 0 — Save plan into the repo
• Step 1 — H1: a stalled write must not delay the teardown
• Step 2 — H2: a source error must not reach go-libvirt's stream sender
• Step 3 — H3: an unreadable address must not look like a missing one
• Step 4 — H4: the base image cache must not be shared across provider entries

Remaining after those: the 48 mediums and lows in review/REPORT.md. Each gets its own checklist line here when it is planned, so this list stays a record of decided work rather than a wish.

Step 0 — plan committed

Copied out of ~/.claude/plans/ and committed alone before touching code, as the repo's plan rules require.

Step 1 — H1: close the SSH client, and signal dead before killing

Two edits, both small, in the order the bug unfolds.

sshSocket.Close was errors.Join(s.Conn.Close(), s.client.Close()). Go evaluates those arguments left to right, so the blocking close ran before the unblocking one. s.Conn is the forwarded libvirt socket, and closing an SSH channel writes a close message under ch.writeMu, which go-libvirt already holds while streaming a base image through the same channel. Against a host that vanished without RST or FIN, that write blocks until the kernel stops retransmitting — tcp_retries2, about 15 minutes. s.client.Close(), a bare net.Conn.Close() on the TCP socket that takes no lock and unblocks the stalled writer through netpoll, was never reached. It now closes the client only.

Dropping the channel close rather than reordering it was the judgment call. Reordering works, but afterwards the channel close returns an error — net.ErrClosed, a *net.OpError, the transport's sticky write error, or io.EOF once mux.loop has run ch.close() — and closeErr is memoised, so a filter that missed a case would poison every later Close and break ssh_test.go:66,69, which assert both closes return nil. Nothing wants the polite channel-close message: dialSSHSocket builds the client solely for this one channel, the sshSocket owns it exclusively, and dialSSHSocket's own error paths already tear down by closing the client.

link.markDead then compounded it by running kill() before close(l.closed), and doing both inside a sync.Once. Reordering alone would not have been enough: sync.Once.Do blocks concurrent callers until the first body returns, so every other markDead — including the one in call's own ctx-cancel branch — queued behind the stalled kill. Cancelling the reconcile did not help. The once field is gone; the mutex the struct already had now guards the "close it exactly once" invariant, reason is set before the signal so a released caller always reads a populated deadReason(), and kill runs outside the lock.

A second route to the same pin, found while planning and fixed by the same edits. go-libvirt's Disconnect closes the net.Conn it was handed (libvirt.go:331 → socket.go:164) — and that conn is our *sshSocket. So libvirtBackend.Close's Disconnect goroutine entered sshSocket.closeOnce and hung there; the 5 s closeBudget released Close, which called markDead("closed") → kill() → sshSocket.Close() → blocked in the other goroutine's sync.Once. The budget did not protect it.

The precondition was verified in the pinned x/crypto v0.54.0 rather than assumed: the flow-control park (remoteWin.reserve, channel.go:268) happens outside writeMu, and only writePacket holds it across the TCP write. So a writer parked in SSH flow control never blocks the close; a writer inside the TCP write always does. Since the channel window is 2 MiB (channelWindowSize = 64 * channelMaxPacket), reproducing it by hand needs a kernel send buffer that fills first — shrink SO_SNDBUF well under 2 MiB.

The older kvm provider does not have this bug: it closes client only (internal/provider/kvm/conn.go:429,456) and never touches the forwarded channel. kvm2's rewrite introduced it.

Tests: planned four, shipped one

The first draft had three hermetic tests plus a //go:build integration && linux stalled-write test with a TCP proxy and shrunken socket buffers. The user pushed back — once the bug is fixed the connection closes properly, so what is left to test? — and they were right about most of it. TestDial_forwardsUnixSocket already covers "Close tears the connection down", and after the fix Close is a single call with no ordering to get wrong. The stalled-write test was the worst of the four: ~150 lines of Linux-only, buffer-autotuning-dependent scaffolding whose own false-green guard made it skip whenever the kernel did not cooperate, to demonstrate a bug that reading x/crypto settles and that the fix deletes the code path for.

What survived is TestMarkDead_slowKillDoesNotBlockCallers, because the markDead change adds a guarantee that did not exist before and nothing else touches it — TestCall_killedByAnotherCallerFailsTransient passes a nil kill, so it cannot see the property. It is also the plausible regression: restoring the sync.Once looks like a correctness improvement. The call's own fn signals when it is in flight, so the parked caller is genuinely parked rather than refused by the dead check on the way in.

The step that actually validates a regression test is running it against the unfixed code. markDead was reverted in place, the test run, and the fixed version restored from a copy:

cp internal/provider/kvm2/backend.go "$SCRATCH/backend.go.fixed"
# ... revert markDead to the sync.Once form and re-add the once field ...
go test -race -count=1 -run 'TestMarkDead_slowKillDoesNotBlockCallers' ./internal/provider/kvm2/
#   --- FAIL: TestMarkDead_slowKillDoesNotBlockCallers (1.00s)
#       backend_test.go:186: link not signalled dead while its kill is still running
cp "$SCRATCH/backend.go.fixed" internal/provider/kvm2/backend.go

It fails in a second with the right diagnostic rather than hanging the suite, because the blocked kill is released by t.Cleanup.

go build ./... && go vet ./... && go test -race -count=1 ./internal/provider/kvm2/

Worth noting

An accepted risk, deliberately not covered by a test. Nothing now stops someone re-adding s.Conn.Close() to sshSocket.Close as tidiness; it would look harmless and every test would still pass, because in a healthy test the channel close returns instantly. The guard is the doc comment on sshSocket, which says outright that closing the channel would queue behind a stalled transfer. That is in idiom for this package — hosts.go:25-34 and create.go:151-153 carry the same kind of hard-won decision — but it is a comment, not a test, and the failure it prevents is a 15-minute operator-wide stall.

Left for later, on purpose. ReconciliationTimeout in SetupWithManager does not exist in this repo at all; adding a per-reconcile deadline would change the contract for every provider, since hosts.go and image.go both cite "no caller deadline", so it needs its own change and its own spec entry. sshSocket.keepalive still has no test (sweep finding S1.2).

Docs. docs/kvm2-provider.md said the adapter "kills the connection (socket and SSH client)", which is no longer true and was the description that made the bug look correct. It now states the order (mark dead, then kill) and why both steps are lock-free. docs/kvm2-provider-spec.md entries 18 and 19 were checked and are not falsified.

No CHANGELOG entry yet — the repo's rule is to add one once the user confirms the change works, so it lands with a later "validate and record" step, as in the kvm campaign.

Step 2 — H2: a source error must not reach go-libvirt's stream sender

This one is a workaround, not a repair. The defect is upstream, and the entry says so in three places (code comment, docs/kvm2-provider.md Known gaps, commit message) so nobody later reads the wrapper as ours to delete.

requestStream in the pinned go-libvirt starts its stream sender on two unbuffered channels and then, on the error path, returns without ever reading outErr. A non-EOF error from the reader makes SendStream emit a StatusError packet and return, whereupon it parks on outErr <- forever; libvirtd answers the StatusError with an error, so the main goroutine parks on abort <- true forever. StorageVolUpload never returns. The 20-minute budget rescues only the caller, and the socket kill does nothing at all here, because both ends are blocked on channel sends rather than on I/O — which is what the comment at backend.go:152 got wrong. Every failed upload leaked two goroutines and 4 MiB against a 128Mi pod limit, with the reconciler retrying on backoff.

Three changes.

eofReader sits at the go-libvirt seam in backend.go, not in image.go, so it also covers the seed and marker uploads, and so it stays out of the file Step 4 rewrites. It converts any source error to io.EOF — the one outcome that ends a stream the normal way — and keeps the real error for StorageVolUpload to return afterwards. countingReader is untouched underneath it and still sees the true error, so uploadImage's existing length and digest checks produce exactly the messages they did before.

The download now carries its own deadline, one minute inside the transfer budget. fetchImage already bounded the whole fetch at 20 minutes, so the body's death and the budget's socket kill were racing, and the socket usually won. With the body a minute ahead, a stalled mirror ends through the reader and unwinds cleanly. This is the difference between a slow mirror costing a Transient error and costing a permanent leak, and slow mirrors are the realistic repeat offender.

discardVolume moved from uploadImage up into fetchImage's upload-error branch, so the dedicated connection is closed — and libvirtd has let go of the stream's fd — before the file is unlinked. Moving it rather than adding a second Close is what keeps image_test.go's closeCount() == 1 assertion true. uploadImage lost its be parameter and its private logger in the process.

What is left, deliberately. A failure that originates in libvirtd mid-upload (ENOSPC, or the volume deleted underneath) still leaks, because by then SendStream has already returned and nothing we pass, cancel or close can reach it. Confirmed unfixable from this side. The upstream fix is two characters and is quoted in Known gaps along with the replace escape hatch, should it ever bite hard enough to be worth the maintenance.

Unverified, ~85%. That libvirtd answers a short-but-clean stream finish with OK rather than an error. virFDStreamWrite enforces length only on writes past it and nothing compares offset to length on the close path, which is also why virsh vol-upload of a small file into a large volume works. If it answers with an error we land back in a deadlock by a different door — but that is precisely today's behaviour, so the change cannot make things worse. To settle it, on any box with a libvirtd (not beelink, which is read-only here), against the pinned go-libvirt:

virsh -c qemu:///session pool-define-as probe dir --target ~/vmprobe
virsh -c qemu:///session pool-start probe
virsh -c qemu:///session vol-create-as probe short.raw 10M --format raw
# then a ~40-line main.go: StorageVolUpload(vol, r, 0, 10<<20, 0) where r
# yields 1 MiB then io.EOF; print the error, then runtime.NumGoroutine()
# and a goroutine profile after a two-second settle.

Expected: err == nil and the goroutine count back at baseline. The same program with a reader that returns a non-EOF error should wedge against the unpatched seam, which also confirms the diagnosis end to end.

The negative check that matters was run on the unit test instead — eofReader.Read was changed back to returning the real error, and both new tests failed with io.Copy through eofReader = connection reset by peer, want nil, then passed again once restored.

go build ./... && go vet ./... && go test -race -count=1 ./internal/provider/kvm2/

Worth noting

The fetchImage discard is still unfenced against a second provider entry that shares the volume name — Step 4 (H4) is what closes that, by making the name per-entry. The deferred atomic-publish work then lands exactly where this step put the discard.

Deliberately not tested: the abort protocol end to end. libvirttest reads a fixed 28-byte header per packet and implements no stream procedures, so pinning this hermetically would mean a 150-250 line libvirt RPC server stub whose only assertion is about upstream's goroutines. Same call as H1's: the test would cost more than it protects. The two tests that did land pin the one property the seam actually needs.

Step 3 — H3: an unreadable address must not look like a missing one

address() tried the guest agent then the host's ARP table and skipped any error that was not a context error, returning ("", nil) after both. An active domain with no IP maps to StateProvisioning, the controller clears status.IP on that, and the health engine prunes the proxy's counters and calls ForgetProxy in one tick once EffectiveHost() is empty.

The damaging trigger was not the guest. The libvirt connection is pooled per host, so when another reconcile's cancel or budget overrun killed it, call returned "connection is closed" — carrying no context sentinel, so the check did not recognise it — and Get returned a successful instance in Provisioning with an empty IP. An unrelated proxy's timeout dropped a healthy proxy out of rotation.

The fix rests on a discriminator the spec had already recorded without drawing the conclusion: during a normal boot on beelink exactly one source fails. The agent errors because it is not up yet, while ARP answers without error and with nothing, since the host never sees the guest's traffic. So one source failing is "no address yet"; every source failing, or a connection already dead, is ErrTransient.

The explicit provider.ErrTransient wrap on the joined error is load-bearing rather than decoration. Without it classify picks the first libvirt code out of the join tree, and ErrInvalidArg, ErrAccessDenied and ErrNoSupport are Permanent — an address read failing would have latched a running proxy as Failed, a worse version of the harm being fixed.

isDead(be) is checked first in the error branch. It is not strictly needed, since a dead link fails both sources and the count rule catches it anyway, but it fails on the first source instead of the second, attributes the error better, and is the same predicate M15 will want for ListByTag.

Tests

TestGet_addressErrorIsProvisioning failed against the new rule, which is correct: it fails both sources with ErrAgentUnresponsive, and code 86 is agent-specific — the ARP source cannot produce it. Re-pointed as TestGet_agentFailsAndARPEmptyIsProvisioning, the beelink boot it was really reaching for: agent errors, ARP returns no addresses and no error, result is Provisioning. Nothing covered that before, and it is the case the new rule must not break.

TestGet_everyAddressSourceFailingIsTransient covers the other side, two subtests: both sources erroring, and the pooled connection dying under the read. Against the unfixed code both return a Provisioning instance instead of an error, which is exactly the bug:

# with isDead() and the every-source-failed arm removed from address():
go test -count=1 -run TestGet_everyAddressSourceFailingIsTransient ./internal/provider/kvm2/
#   get_test.go:215: Get = &{... IP: State:Provisioning ...}, want no instance

TestGet_arpFallbackWhenAgentFails passes untouched — one source of two.

Worth noting

Half of H3 is deliberately still open, and it is the half that fires more often. A guest agent restarting inside a running VM still demotes it to Provisioning and clears its IP: the agent errors, ARP answers empty, and that is indistinguishable from a VM still booting. The discriminator is the previous IP, which the provider does not hold by design (TestGet_readsUIDAndCreatedAtFromBackend pins "read back, not remembered") and the controller already has. Follow-up finding, to be filed against internal/controller/proxy_controller.go:208-211: on StateProvisioning, keep an existing IP and Provisioned condition for a few polls rather than clearing on the first, or give the health engine a grace period before pruning.

The one kvm2-local alternative was considered and rejected: the mark carries createdAt, so get could treat "active, no address from any source, domain older than a provisioning window" as an error. It uses state the provider genuinely has, but createdAt is creation rather than boot, it fires on every genuinely stuck VM as a retry-forever error, and it invents a provisioning deadline that belongs to the controller. A design change wearing a bug fix's clothes.

L6 not folded in. A domain that vanishes between DomainGetState and the address read now returns Transient rather than a fake Provisioning — better, and self-correcting on the next poll. Returning NotFound one poll earlier is a two-line addition at the same spot, left for when L6 comes up so this commit stays one idea.

Step 4 — H4: the base image cache must not be shared across provider entries

baseVolumeName was base-<key>.<format> — byte-identical to what internal/provider/kvm/image.go builds in the same pool, for URL images too (kvm takes hex(sum[:8]), kvm2 takes hex(sum[:])[:16], the same eight bytes), with a completely separate lock table. The shipped config has a live kvm-home entry and a commented-out kvm2-home entry on the same host, the same key debian-13 and the same default pool, and the comment invites the operator to enable it. Both integration tests use key test-image. On beelink both live overlays back onto kvm's base-debian-13.qcow2.

Since cachedImage accepts any 64-hex marker when no sha256 is configured and never re-hashes the base, that meant: two entries pointing one key at different URLs silently share a file; a concurrent cold fetch lets one entry delete the other's half-uploaded base and boot a guest from the remains; and a sha256 edit deletes the base in place under live overlays, which libvirt permits without complaint.

The name now carries the provider entry: base-<entry>-<key>.<format>. imageSource gained an Entry field and resolveImage takes the provider name, so the four production call sites all follow from imageSource.volumeName(). That closes the cross-entry sharing outright, kvm↔kvm2 and kvm2↔kvm2 alike, and it turns the existing per-Provider imageLocks from a latent bug into a correct one — with names unique per entry, no two entries can contend for a file, so a process-wide lock table is not needed.

The entry name reaches a file name, so it now has to be a usable one. Provider names are validated for all types at once and only for being non-empty and unique (internal/provider/config.go), so New checks its own against the existing kvmIdentifier rule, exported as provider.IsKVMIdentifier rather than duplicated. A name with a slash fails at startup instead of producing a volume nothing can address.

One companion fix, because the naming alone does not stop it. cachedImage's digest-mismatch branch deleted the marker and the base and refetched. It now returns ErrPermanent naming both digests and telling the operator what to do. That branch was the remaining path that unlinks a good file under running domains, and the contract already said "add a new key instead" — the code was contradicting it. The no-marker and non-hex-marker branches keep deleting, since there the base is genuine debris.

What this does NOT fix

Stated plainly rather than left to look closed:

• Editing images.<key>.url is still a silent no-op. The cache is keyed by entry and key, and nothing records the URL, so the new one is never fetched. Pre-existing, already documented as "add a new key"; the eviction section now says it outright.
• Cross-process fencing. Two managers, or a manager plus a hand-run virsh, can still meet on one name. E4 refuted the rollout and leader-handover triggers for the shipped manifests, so this is narrow, and its fix is the deferred atomic publish — upload to a temp name, verify, publish — which lands where H2 put the discard.

Migration

Nothing to do. The base-debian-13.qcow2 on beelink is kvm's, kvm keeps using it untouched, and kvm2 downloads its own copy once per host when kvm2-home is enabled. No window in which anything is replaced under a running guest. The pool needs room for both, which the prerequisites now say.

Tests

Mostly literals becoming helpers, plus a new fx.baseName() in provider_test.go for the three fixtures that asserted "base-debian-13.qcow2" by hand. The log and span assertions in TestEnsureImage_downloadsAndWritesMarkerLast keep asserting a full name rather than loosening to a prefix, so they still pin that the key stays readable in operator-facing output.

One test changed behaviour rather than spelling: TestEnsureImage_sha256ChangeRedownloads became TestEnsureImage_sha256ChangeRefusesInsteadOfReplacing, asserting Permanent, zero server hits, zero deletes, and both volumes untouched. Two added: TestNew_rejectsUnsafeEntryName, and a TestVolumeNames row for a second entry with the same key.

Both negative checks were run:

# baseVolumeName reverted to base-<key>.<format>:
go test -count=1 -run TestVolumeNames ./internal/provider/kvm2/
#   volume_test.go:135: got "base-debian-13.qcow2", want "base-kvm2-lab-debian-13.qcow2"

# the digest branch reverted to delete-and-refetch:
go test -count=1 -run TestEnsureImage_sha256ChangeRefuses ./internal/provider/kvm2/
#   image_test.go:412: ensureImage() error = <nil>, want ErrPermanent telling the operator what to do

go vet -tags=integration was run too, since kvm2_integration_test.go:484 builds the base name and only compiles under that tag.

Worth noting

The integration test has still never run on beelink. Step 11 is unchecked in docs/plans-executions/2026-09-16-1333-kvm2-provider.md, every base-* on the host predates kvm2, and both live domains carry the kvm v1 metadata namespace. After this change it is also the first end-to-end exercise of the new name, and it will do a real cold download under base-kvm2-integration-test-image.qcow2 beside kvm's untouched base-test-image.qcow2.

Docs touched beyond the obvious: the eviction recipe named base-debian-13.qcow2, which after this change is kvm's file — following it would have deleted the wrong provider's base. It now greps for the prefix and deletes the marker first.