CVSROOT: /cvs Module name: src Changes by: d...@cvs.openbsd.org 2023/11/09 11:36:19
Modified files: sbin/pflogd : pflogd.c Log message: avoid restartable syscalls with siginterrupt() against all our handlers. pflogd uses blocking bpf reads, but installs a bunch of signal handlers to handle cleanly closing and (re)opening the log file. signal() by default sets the handlers up so they're restartable. this has the effect that when pflogd receives a signal while waiting in bpfread, the signal handler runs and sets a flag saying the file should be rotated or closed or whatever, but then the kernel restarts the read. when pflogd used a bpf read timeout, pflogd would run it's "event" loop every time the read timeout expired. this meant even though the bpf read was restarted, by having the read timeout expire every 500ms the flag the signal handlers set would be processed in a relatively short period of time. after moving bpf to a wait timeout, pflogd basically has to wait for a packet to be captured before the bpf read will return. if you're trying to kill pflogd on an idle firewall then you're basically stuck. making the signal handlers not restartable allows bpfread to fail with EINTR so pflogd can go around it's even loop and exit as expected. reported by Mikhail on bugs@ ok claudio@ pflogd needs a rewrite though.