[BUG] kernel freeze, rcu_sched self-detected stall on CPU
From: Johanna Abrahamsson
Date: Sat Dec 17 2016 - 17:26:08 EST
A bug has been introduced in ipc/sem.c by a patch '[PATCH -next 2/2] ipc/sem: avoid idr tree lookup for interrupted semop' (https://marc.info/?l=linux-kernel&m=147870885903619, commit 370b262c896e5565b271a3ea3abee4d0914ba443).
The bug results in a kernel freeze and several "INFO: rcu_sched self-detected stall on CPU". The system becomes unresponsible and has to be rebooted to work again as far as I can tell.
My repro case is, hilariously enough, packaging the linux kernel for debian with `make -j8 deb-pkg`. I'm sure a smaller repro case could be constructed, but so far I've had 100% reproducibility from this one. The freeze always occurs at the Makefiles.headerinst part of packaging, after the kernel is built.
The bug is present in next-20161124 and linux-next versions going forward from that and has now been merged to linus/master. However, in linus/master it doesn't trigger a bug. This seems to indicate that it only triggers a bug in combination with a commit that is present in next-20161216 but not present in linus/master.
I've confirmed that doing `git revert 370b262c896e5565b271a3ea3abee4d0914ba443` on next-20161216 fixes the bug.
In all repro cases I've tested on virtual machines with multiple cpu cores running under either Xen or Virtualbox.
I will try to investigate this further but as I have limited knowledge of RCU and how the kernel works with semaphores don't expect any miracles :)
Best Regards,
Johanna Abrahamsson