Skip to content

v1.6.0

Choose a tag to compare

@heyvaldemar heyvaldemar released this 14 Nov 10:04
· 5 commits to main since this release

Added

  • The timer reports a crash Docker restarted. The health check cannot see one: it runs inside the container, which the restart replaces. tools/steam-uptodate.sh now writes down the container's name, id and Docker's RestartCount on each run and, when the count grows, says once "the server restarted N times since , last exit code X, last finished " with the signal named and the reason when it can tell (Docker's oom event, or the last crash line from the log, long tokens masked); three or more in 24 hours exit 1. A recreated container and a count set back to 0 by docker restart start afresh (so a crash in the same window as a restart by hand is not seen). The exit code and a kill for memory come from Docker's die and oom events: a restarted container's .State.ExitCode reads 0 and .State.OOMKilled false again. Proven against real containers under the restart policy, one that crashes and one that runs out of 16 MB (tests/restarts.sh) and on the real server (the game killed with SIGSEGV).

Fixed

  • A server with only bots on it no longer crashes after some 15,000 bot deaths. weapon_max_before_cleanup and weapon_auto_cleanup_time default to 0, so every dropped weapon stayed on the map. With only bots on, the warmup waits for a human and never ends, each bot death left a pistol behind, and at 16,384 networkable entities the engine stopped ("Ran out of networkable entities") and segfaulted; Docker restarted it with the health check green. On servers I run with real players that happened once on casual maps and eleven times on classic ones since August. hooks/pre.sh now writes 64 and 120 into server.cfg, which runs on every map, a workshop map's included, as one marked block it replaces on every start, so the lines never pile up; two variables change them. Real Server reads both over rcon after the start and after a map change.

Changed

  • The secret probe watches the rcon client too. tests/secret-probe.sh during samples the host's ps and every container's command lines the whole time a helper runs; Real Server wraps the rcon map change in it, so a password handed to the client as an argument, which lives only while it runs, would be caught.

Verified before release

  • Real Server 43580304594 on 8771c05, a real CS2 with a real game server token: weapon_max_before_cleanup 64 and weapon_auto_cleanup_time 120 read over rcon after the start and after a map change; after a restart, exec secrets.cfg and both weapon lines in server.cfg once each; the rcon client watched by the secret probe the whole time it ran, no value on any command line; the game killed with SIGSEGV, Docker restarted it, and the timer said so once, with a crash line, and stayed silent on the next run; the probe clean after the start and after a restart.
  • Deployment Verification 43580304638: tests/restarts.sh against real containers under the restart policy (a crash said once with its exit code and line, a restart by hand starting the count afresh, three in 24 hours exiting 1, a 16 MB container killed for memory named as such), and tests/steam-uptodate.sh, 41 cases.