The bug was somehow masked for a long time and exposed somewhere between 6.0 and 6.1.
At 33c3 I noticed that my old and weathered i386 netbook had stopped resuming. Turns out
the kernel crashed during resume on a secondary CPU when it tried to run the first bit of userland code there
(which happened to be syslogd) because (as far as I understand) some registers describing address space mappings were not restored. It took a long time for me to get any information on this bug
because with an ssd it resumed so fast that the system crashed before the screen came on (and no serial console…).
A week ago I swapped a slow spinning disk in. Luckily this helped and I finally saw what ddb was writing to the screen.
I am still in awe of how kettenis figured this out within 30 minutes of seeing my report.
Don’t forget about the resume fix for multi-core i386: http://marc.info/?l=openbsd-cvs&m=149072212601001&w=2
The bug was somehow masked for a long time and exposed somewhere between 6.0 and 6.1. At 33c3 I noticed that my old and weathered i386 netbook had stopped resuming. Turns out the kernel crashed during resume on a secondary CPU when it tried to run the first bit of userland code there (which happened to be syslogd) because (as far as I understand) some registers describing address space mappings were not restored. It took a long time for me to get any information on this bug because with an ssd it resumed so fast that the system crashed before the screen came on (and no serial console…). A week ago I swapped a slow spinning disk in. Luckily this helped and I finally saw what ddb was writing to the screen.
I am still in awe of how kettenis figured this out within 30 minutes of seeing my report.