[kolyshkin]

The --timeout timer is set with g_timeout_add_seconds, which rounds the expiration time to a whole second boundary (adjusted by a per-host offset) to group wakeups. When the sub-second remainder is less than 250ms it is rounded down, so the timer can fire up to 250ms before the requested timeout (see g_timeout_set_expiration in glib's gmain.c).

As a result, podman run --timeout 2 IMAGE sleep 60 sometimes takes less than 2 seconds in total, and podman's podman run --timeout - basic test system test flakes with:

FAIL: podman stop: ran too quickly!
expected: -gt 1
  actual:     1

(seen in runc CI: https://github.com/opencontainers/runc/actions/runs/40177692604/job/121122129191, where the whole podman run took 1.92s).

Use g_timeout_add with milliseconds instead, which does not round. Keep g_timeout_add_seconds for timeouts that do not fit into guint milliseconds (about 49 days), as the imprecision does not matter there.

Measured locally (rootless podman 5.8.9, 12 runs of podman run --rm --timeout 2 ... sleep 60, container setup takes ~100ms):

conmon	total time, ms
2.2.1	1976–2019
this PR	2103–2123

[jnovy]

LGTM, nice fix.

The overflow guard (guint)opt_timeout <= G_MAXUINT / 1000 is correct - max product is 4294967000 which fits in guint. The (guint) cast is safe because the outer opt_timeout > 0 check guarantees a positive value (opt_timeout is int per cli.c:42). Fallback to g_timeout_add_seconds() for >49-day timeouts is a sensible choice since 250ms imprecision is irrelevant at that scale.

Builds clean with -Werror locally, no warnings.

One note: the PR branch needs a rebase onto current main - the diff includes stale GitHub Actions version pins from before renovate PRs #700, #701, #704 merged. Without rebase, merging would revert those updates.