Hi Finn,
On 10/09/21 10:51, Finn Thain wrote:
On Wed, 8 Sep 2021, Michael Schmitz wrote:
In a related case, I've managed to swap my 'resume_userspace' format
error for a nice 'illegal instruction' format error apparently caused by
an invalid function pointer in __handle_irq_event_percpu(), just by
disabling all interrupts upon entering the auto_inthandler and
user_inthandler exception handlers. This bug is quite readily reproduced
by running your kernel_coverage.sh script in a loop (panics on the first
stress test on the second pass):
Stress run 2
Logging to stress-ng-20210908-0838.log
./kernel-coverage.sh: line 272: lcov: command not found
running --fork 1 --fork-vm -t 60 --timestamp --no-rand-seed --times
stress-ng: 08:40:08.70 info: [1914] setting to a 60 second run per stressor
stress-ng: 08:40:08.82 info: [1914] dispatching hogs: 1 fork
packet_write_wait: Connection to 10.1.1.4 port 22: Broken pipe
Why disabling interrupts during interrupt processing would make matters
worse doesn't make any sense to me...
Are you able to reproduce that with a stock mainline build?
I still need to try that. The kernel panic went away when adding
additional local_irq_save()/local_irq_restore() in do_IRQ(), which
should done have nothing except slightly increase code size. Now trying
to find out whether interrupts are reenabled any time during interrupt
processing ...
Cheers,
Michael
