vms-707: DCL RUN surfaces the image's executive-recorded $STATUS (not 0x1)

[baron-3dl]

Problem (vms-707, p1 — critical path)

DCL's RUN verb forks a subprocess to activate an image and read its exit, but the fork()+execve() fallback in dcl_activate_image collapsed the POSIX child exit to SS\$_NORMAL/SS\$_ABORT, discarding the image's real VMS condition value. An image whose main returned 3 reported \$STATUS %X00000001 instead of the faithful DEC C encoding (C\$_EXIT1 + (3-1)*8 = 0x0035A019) the executive recorded at SYS\$EXIT. This blocked reading any fork-activated image's true exit status (e.g. the GCC-lane #846 crtl_rms sentinel).

The in-process activation path already read the executive-recorded \$STATUS (vms_kif_getexit); only the fork fallback collapsed it.

Fix — RUN reads the executive-recorded condition value

VMS behavior: the image records its completion condition value and the invoking CLI reads it back as \$STATUS. OVMX's fork child records its condition into its own executive PCB at SYS\$EXIT. Because that child shares DCL's VMS PID (REGISTER_CONTINUE), a by-VMS-PID read is ambiguous between DCL and the child — so RUN reads by the child's Linux pid, which names its PCB row uniquely, in the zombie window (WNOWAIT peek) before waitpid reaps it. Falls back to the POSIX-derived verdict only when no image recorded a status (no /dev/vms, or a foreign tool that never \$EXITs).

• src/kernel/vms_ioctl.h, src/kernel-netbsd/vms_proctab_nb.h: VMS_JPI_SEL_LINUX_PID (no struct/ioctl-number change; reuses the args' vms_pid field).
• src/kernel-core/vms_proctab.c: find_by_linux_pid + the SEL_LINUX_PID GETEXIT case (shared by the Linux vms.ko and the NetBSD module).
• src/libvmssys/vms_kif.{c,h} + src/vmslink/libvmssys_shr.vec: vms_kif_getexit_linux wrapper (census-satisfied: DCL calls it).
• src/vmsdcl/dcl_cmd_process.c: RUN fork path reads the recorded \$STATUS.

stdout/stderr routing

Investigated: tests/qemu/test_product_install_e2e.sh already RUNs an installed image via the fork path and asserts real console output, so fork-path stdout/stderr routing to the console already works. The BOOT-A "absent output" was the opt-in OVMX_IMGACT_SEAM line (off by default) plus that image's own path — not a routing regression. No routing change needed; the \$STATUS collapse was the substantive bug.

Proof / regression gate

tests/qemu/test_kmod_exit.c Part 5 (kernel-executive QEMU suite, real /dev/vms): a REGISTER_CONTINUE child records 0x0035A019, the parent confirms it shares DCL's VMS PID, then reads the child's \$STATUS by Linux pid in the zombie window — the exact dcl_activate_image sequence — and asserts the EXACT condition round-trips (not 0x1), plus a negctl that an unbacked Linux pid reads SS\$_NONEXPR.

The \$STATUS symbol chain is verified end-to-end: cmd_run → dcl_activate_image returns cond → dcl_set_status renders %X0035A019 into the \$STATUS symbol read by SHOW SYMBOL \$STATUS.

🤖 Generated with Claude Code