kill: implementation for Windows

[nikolalukovic]

Implement kill util for Windows since #13363 added most of plumbing needed.

[codspeed-hq]

Merging this PR will degrade performance by 2.15%

⚠️ Different runtime environments detected

Some benchmarks with significant performance changes were compared across different runtime environments,
which may affect the accuracy of the results.

Open the report in CodSpeed to investigate

⚡ 6 improved benchmarks
❌ 14 regressed benchmarks
✅ 249 untouched benchmarks
⏩ 122 skipped benchmarks¹

Warning

Please fix the performance issues or acknowledge them on CodSpeed.

Performance Changes

	Mode	Benchmark	BASE	HEAD	Efficiency
❌	Simulation	hostname_ip_lookup[100000]	250 µs	275.6 µs	-9.32%
❌	Memory	hostname_basic	56.3 KB	59.8 KB	-5.9%
❌	Memory	seq_custom_separator	58.4 KB	61.9 KB	-5.7%
❌	Memory	nl_many_lines[100000]	66.2 KB	69.7 KB	-5.06%
❌	Memory	expand_custom_tabstops[50000]	67 KB	70.5 KB	-5%
❌	Memory	cut_fields_newline_delim	69.3 KB	72.8 KB	-4.85%
❌	Memory	fold_custom_width[50000]	76.1 KB	79.6 KB	-4.43%
❌	Simulation	ls_recursive_balanced_tree[(6, 4, 15)]	112.5 ms	117.5 ms	-4.29%
❌	Memory	split_numeric_suffix	80.5 KB	84 KB	-4.2%
❌	Memory	df_deep_directory	80.7 KB	84.3 KB	-4.19%
❌	Memory	join_custom_separator	86 KB	89.5 KB	-3.94%
❌	Memory	cksum_multiple_files	91.2 KB	94.7 KB	-3.73%
❌	Memory	rm_recursive_tree	98 KB	101.5 KB	-3.48%
❌	Simulation	single_date_now	191.5 µs	197.7 µs	-3.1%
⚡	Simulation	ls_recursive_deep_tree[(200, 2)]	3.6 ms	3.4 ms	+6.05%
⚡	Simulation	cksum_crc32b	41.9 ms	39.7 ms	+5.71%
⚡	Simulation	numfmt_from_si[10000]	84.9 ms	81.4 ms	+4.32%
⚡	Simulation	ls_recursive_mixed_tree	2.9 ms	2.8 ms	+3.88%
⚡	Simulation	cksum_default	40.4 ms	39.2 ms	+3.06%
⚡	Simulation	cksum_crc	41 ms	39.8 ms	+3.01%

Tip

Investigate this regression by commenting @codspeedbot fix this regression on this PR, or directly use the CodSpeed MCP with your agent.

---

_{Comparing nikolalukovic:feature/kill-windows (c2377ca) with main (eae5c43)}

Footnotes

1. 122 benchmarks were skipped, so the baseline results were used instead. If they were deleted from the codebase, click here and archive them to remove them from the performance reports. ↩

[github-actions]

GNU testsuite comparison:

Skipping an intermittent issue tests/cut/bounded-memory (passes in this run but fails in the 'main' branch)

[lhecker]

Why not use TerminateJobObject here?

[nikolalukovic]

how do we get the handle?

[lhecker]

There's no direct public Win32 or NT API for that. So, we got to get creative, and creative is unfortunately a bit involved. If you're using AI to code this, I assume it won't trouble you too much. Otherwise, let's consider this a TODO. The idea would be to iterate through all processes using NT APIs and filter down to the current process and get its Job ID. Then get all handles with SystemExtendedHandleInformation and get the matching Job handle and kill it. Something like that - it's probably quite a bit of code. (More than I would've assumed intuitively. 🙁)

[nikolalukovic]

lets give it a shot

[nikolalukovic]

kill 0 — the four implementations

	Unix (POSIX)	This Windows port	Cygwin / MSYS2	busybox-w32
Target set	Every process whose pgid == caller's pgid	Every process in the caller's immediate Job object and its child jobs	Every Cygwin process with pgid == myself->pgid && ctty == myself->ctty	Caller + its descendants (parent-pid tree walk)
Grouping primitive	Kernel process group (setpgid/setsid)	Kernel Job object	Userspace _pinfo table in Cygwin's shared memory	None — recomputed per call from a snapshot
Who created the group	The caller (or its shell), deliberately, to bound a blast radius	Someone else — terminal, IDE, Docker, CI agent, PCA — as a kill switch for you	Cygwin's own runtime	N/A
Enumeration	Kernel does it; no snapshot	QueryInformationJobObject(NULL, JobObjectBasicProcessIdList) — snapshot, so a process spawned mid-sweep escapes	Cygwin's private NT object directory	CreateToolhelp32Snapshot — snapshot, same race
Delivery	Real signal	Per-PID OpenProcess + TerminateProcess, caller last	Real emulated signal via the target's signal pipe; its signal thread runs the handler	TerminateProcess per pid
Catchable / blockable / ignorable	Yes (except KILL, STOP)	No — nothing is catchable	Yes, genuinely	No
Reaches non-participating processes	All, subject to permission	Yes — any native process the job holds	No — a native process unrelated to Cygwin has no _pinfo entry and is unreachable	Yes, if it's a descendant
Caller signalled too	Yes	Yes, deliberately last so it doesn't die mid-sweep	Yes	Yes

Concrete invocations

Invocation	Unix	This Windows port	Cygwin	busybox-w32
kill -0 0	Probe; 0 if any member signallable	Probe every member (OpenProcess + 0-timeout wait); terminates nothing	Probe	Probe
kill 0	SIGTERM, handlers run	Force-terminate all, exit 143 each	SIGTERM, handlers run	Force-terminate
kill -9 0	SIGKILL	Force-terminate, exit 137 each	SIGKILL	Force-terminate
kill -INT 0	SIGINT, handlers run	Force-terminate, exit 130 — not a CTRL_BREAK (that path exists but only for a console group led by a known pid)	SIGINT, handlers run	Force-terminate
kill -STOP 0	Suspends the group	Rejected, unsupported signal on Windows	Suspends (SuspendThread on all threads)	Rejected
kill -TSTP 0	Suspends (unless caught)	Silent no-op (Disposition::Ignore) — legitimate, TSTP is catchable	Suspends	—
kill -TERM -N	Signals group N	Rejected, negative-PID error	Works, any pgid	—
Caller in no group/job	N/A — always in a group	Signals only itself	N/A	Self only, if childless
Some members denied	0 if ≥1 succeeded	Same: Ok if ≥1 delivered, first error if none	Same	—
Exit status of a member	WIFSIGNALED, shell shows 128+n	Literal exit code 128+n — WIFSIGNALED is not representable	WIFSIGNALED (emulated properly)	TerminateProcess code