mirror of
https://github.com/Klipper4FlashForge/firmware.git
synced 2026-10-21 21:02:16 -05:00
Oleksandr ShyshatskyiandClaude Opus 5
39710c0f02
s6-rc-init and s6-rc default to TAIN_INFINITE_RELATIVE. skalibs converts a deadline to a struct timeval before select(), and infinity does not fit a 32-bit time_t: the call fails immediately with EOVERFLOW -- "Value too large for defined data type" -- before a single service is touched. Passing -t fixes it, in firmwareExe, start.sh and klipperDaemon. WHY IT APPEARED NOW. qa-suite-bridgehead built the supervision stack with the Bootlin musl toolchain, and musl has a 64-bit time_t on every architecture. pkgs/lib.sh states the rule this repo actually follows -- "ONE TOOLCHAIN: everything here cross-compiles with the Ingenic glibc 2.29 / gcc 7.2 toolchain" -- and glibc 2.29 on mips32 does not. So the bug was latent in the graph and became visible the moment s6 became a recipe like everything else. Fixing it in the callers keeps one libc, which is the point of that rule. HOW IT WAS FOUND, because three plausible answers were wrong first. Not the filesystem: it failed identically with the live dir on /usr/data and on /run. Not the database: s6-rc-db read it fine, and a database recompiled inside the replica by the target compiler failed the same way. Not a missing -D_FILE_OFFSET_BITS=64: both s6-svscan and s6-rc-init reference readdir64 and never plain readdir -- and s6-svscan supervises a NON-empty scandir on both filesystems, which is the test that had been missing (it had only ever been started on an empty one). qemu's own strace then showed no failing syscall at all, which is what pointed at a deadline computed in userspace. The boot now reaches the transition: the scanner comes up, s6-rc-init lays down all seven servicedirs and links them, and s6-rc starts wifi, nginx and moonraker. qa/replica/test_s6rc.py still fails, for an environmental reason that is not this bug: the replica creates null, zero, full, random, urandom, tty and fb0 and no ttyS*, so mcu-bringup cannot reach the toolhead boards, klipper never starts, and the ok-all bundle never comes up. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01VpUmBb4xtmnmtTosonqeNi