Three audit claims traced: a BAR asked for twice panics the kernel, netstack keeps a dead client's socket, and a PCI call's slot is not its claim

[Japabu]

An outside static audit of main at b432ed21c claimed three defects in the PCI claim substrate and in netstack. Each was argued from reading and run by nobody. This traces each in the tree, runs the two a test can reach, and records all three. It fixes none: the diff against main is three files under issues/.

Head 23f36bce9, on main 017451624.

Verdicts

claim	verdict	issue
a later BAR request panics on the cached retired object	REPRODUCED	issues/a-claims-bar-asked-for-again-after-its-handle-closed-panics-the-kernel.md
netstack keeps the TCP socket of a finished bridge	REPRODUCED	issues/netstack-keeps-the-socket-of-a-connection-whose-client-sent-no-close.md
a PCI call keeps only a reusable slot number across its claim's release	NOT REPRODUCED: not run, and no step refuted	issues/a-pci-call-in-flight-names-its-function-by-a-slot-the-next-claim-reuses.md

The two red tests, and where they are

The tree carries no known-red test: a red test is deleted, and its issue names the commit that restores it. So each test is one commit of this branch and its deletion another:

test	added	deleted	restore
bar_map_again	ef41f0ad1	a86cafa06	git revert a86cafa06
netstack_socket_churn	06bbc236d	8dd55d06a	git revert 8dd55d06a

Each revert was applied alone on the tip and applies clean.

Both were run at 06bbc236d, the commit that carries both:

• cargo test --test toyos-build -- bar_map_again: exit 1.

  [ 2.609 cpu0 kernel alert] PANIC: panicked at src/object/handle.rs:108:9:
  a handle to a retired SharedMem (koid 197)
  

  after the job's bar_map_again: BAR 4 asked for and its handle closed; asking again.
• cargo test --test toyos-build -- netstack_socket_churn: exit 1.

  netstack_socket_churn: 4 connections ended and let go; netstack held 0 stream(s) before them and holds 4 after
  

• The control for the second, the same test with the client's close request sent before it drops its pipes (patch in the first comment): exit 0. The red is the missing request and nothing else about the run.
• cargo test --test toyos-checks with both tests in the tree: exit 0, 36 passed.

There is no control for the first short of a fix: its green arm is a kernel that answers.

Why these are guest tests

• bar_map_again: the path is pcidev::bar_object into ops::install, both in kernel/src, which has no host build; no type refuses it today. A metal row reaches it through the T14's I219, but its red is that machine in a panic, and the T14 is not this agent's to run.
• netstack_socket_churn: the state is Netstack's in userland/netstack/src/main.rs, a binary whose manifest exempts it from host builds and whose bridge reads kernel pipes; it has no host test to extend. It needs a real stack with a peer that ends connections, which under QEMU is a listener on the host's loopback.

The one not run

Every step of the third claim holds on reading: pci_slot answers a bare slot number and lets go of the claim; with_bound checks that the slot is bound and unfaulted, not whose binding it is; the two are under different locks; slot::reserve hands the first free slot to the next claim. Nothing found stops it.

It is not run because a guest cannot order a release and a second claim between two adjacent statements of one syscall. A test that does not depend on luck needs the kernel to hold a call at that point, as the existing SYS_DEBUG actuators hold other windows. That is kernel source, outside a change that touches issues/ and tests only, and a timing race in its place would be a flaky test. The issue's exit names that test.

Gates

• cargo run -- --ci host at 23f36bce9: exit 0, Host: 77 step(s), all green.
• No guest test is reached by the head's diff, which is issues/ alone.

Unsure of

• The second way the first defect's object is retired, a first install refused for a full handle table, is read from the code and not run.
• The audit's report is a local file outside git. Nothing of its text is in the tree or here; the issues are written from the trace and the runs.
• a-pci-call-in-flight… is filed as defect though unrun, as netstack-removes-a-closed-stream-before-its-fin-leaves.md is: an isolation finding is not left as a finding.

🤖 Generated with Claude Code

https://claude.ai/code/session_01RvnWQFcMuGqTHYhvSnTe8A

[Japabu]

The control for netstack_socket_churn, applied on 06bbc236d, run (exit 0) and reversed in one script:

--- a/tests/toyos-rust-tests/src/bin/netstack_socket_churn.rs
+++ b/tests/toyos-rust-tests/src/bin/netstack_socket_churn.rs
@@ -41,6 +41,7 @@
         let mut byte = [0u8; 1];
         assert_eq!(conn.rx.read(&mut byte), Ok(0), "connection {round}: the host ends it without a byte");
         // Both pipe ends, and no close request.
+        toyos::net::tcp_close(conn.socket_id).expect("the control: the close request this test leaves out");
         drop(conn);
         let asked = Instant::now();
         // Each question is a round trip netstack answers on a pass of its own.

[Japabu]

Review, round 1, of 23f36bce9 against .claude/agents/reviewer.md

Net: +123 lines in 3 files, all under issues/; production 0, tests 0. The four test commits cancel: git diff origin/main a86cafa06 is empty.

What was checked

• The reverts. a86cafa06 reversed is patch-identical to ef41f0ad1; 8dd55d06a's parent is 06bbc236d, so its revert brings back the guest binary, both rows, the arm and body, and tests/netcase's receives. Each revert merges clean alone on the head (git merge-tree --merge-base=<commit> <head> <commit>^, exit 0 for both). main lands branches by merge commit, so both commits will be reachable from it.
• The logs. cap01-06bbc236d.log: exit status 1, PANIC: panicked at src/object/handle.rs:108:9: a handle to a retired SharedMem (koid 197), backtrace through ops::install from sys_device_bar_map, pid 10 test_rs_bar_map_again. srv01-06bbc236d.log: exit status 1 on held 0 stream(s) before them and holds 4 after. srv01-control-06bbc236d.log: PASS, with the first comment's one-line patch. ci-host.log: started one second after the head's commit time, Host: 77 step(s), all green, EXIT=0.
• BAR (executed, seen). The tree agrees with the panic: bar_object (kernel/src/pcidev/mod.rs:1501) keeps the object in bound.bars, HandleEntry::drop retires it at zero handles, HandleEntry::new asserts. The second arm, read and marked so, holds: ops::install builds the entry before HandleTable::install can answer TableFull, and the refused entry drops.
• netstack (executed, seen). bridge_piped ends in piped_connections.swap_remove alone (userland/netstack/src/main.rs:1389); handle_tcp_close removes all three (:794); piped_live counts bridges and pending connects; sockets.tcp counts table entries.
• Slot (traced). Every step holds: pci_slot (kernel/src/syscall/device.rs:168) and sys_device_reg's RegTarget::PciConfig answer a bare number and drop the claim; with_bound (pcidev/mod.rs:1488) checks faulted and bound and nothing of whose; the process-data lock and BOUND[slot] are separate with nothing across them; release leaves a reset function's slot Free and slot::reserve gives a requester its own residue, else the first free slot. A thread is preempted at the lock release that ends pci_slot (the Ring 0 tick sets need_resched, preempt::enable acts on it).
• Fields and kind. Three status: open, kind: defect, opened: 2026-10-08; slugs unique; both owner tracks and the cited issue exist; no existing issue records any of the three. defect for the unrun one is right by issues/README.md: what a finding is promoted to is "something real that someone should act on — a fix, a measurement, an instrument", and "when unsure, promote"; main already files netstack-removes-a-closed-stream-before-its-fin-leaves.md so. The change that makes the two branches agree is in Three audited kernel races recorded, traced and unexecuted: an IPC rollback's expect, chdir against open, an shm map against the last close #754. No machine is named beyond what the tree already names; the BAR issue's "ask, close, ask" says no more than its title.

NOTE

• issues/a-claims-bar-asked-for-again-after-its-handle-closed-panics-the-kernel.md:36 and the PR body — the job's line BAR 4 asked for and its handle closed; asking again is credited to the run at 06bbc236d; that run's log has an empty "the job said:", and the line is in cap01-run1.log, a run made before the commit existed — drop the clause or quote the run that shows it.

LAND